/*
* alltoall_intra_linear_sync
*
* Function: Linear implementation of alltoall with limited number
* of outstanding requests.
* Accepts: Same as MPI_Alltoall(), and the maximum number of
* outstanding requests (actual number is 2 * max, since
* we count receive and send requests separately).
* Returns: MPI_SUCCESS or error code
*
* Description: Algorithm is the following:
* 1) post K irecvs, K <= N
* 2) post K isends, K <= N
* 3) while not done
* - wait for any request to complete
* - replace that request by the new one of the same type.
*/
int ompi_coll_base_alltoall_intra_linear_sync(const void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
void* rbuf, int rcount,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module,
int max_outstanding_reqs)
{
int line, error, ri, si, rank, size, nreqs, nrreqs, nsreqs, total_reqs;
char *psnd, *prcv;
ptrdiff_t slb, sext, rlb, rext;
ompi_request_t **reqs = NULL;
if (MPI_IN_PLACE == sbuf) {
return mca_coll_base_alltoall_intra_basic_inplace (rbuf, rcount, rdtype,
comm, module);
}
/* Initialize. */
size = ompi_comm_size(comm);
rank = ompi_comm_rank(comm);
OPAL_OUTPUT((ompi_coll_base_framework.framework_output,
"ompi_coll_base_alltoall_intra_linear_sync rank %d", rank));
error = ompi_datatype_get_extent(sdtype, &slb, &sext);
if (OMPI_SUCCESS != error) {
return error;
}
sext *= scount;
error = ompi_datatype_get_extent(rdtype, &rlb, &rext);
if (OMPI_SUCCESS != error) {
return error;
}
rext *= rcount;
/* simple optimization */
psnd = ((char *) sbuf) + (ptrdiff_t)rank * sext;
prcv = ((char *) rbuf) + (ptrdiff_t)rank * rext;
error = ompi_datatype_sndrcv(psnd, scount, sdtype, prcv, rcount, rdtype);
if (MPI_SUCCESS != error) {
return error;
}
/* If only one process, we're done. */
if (1 == size) {
return MPI_SUCCESS;
}
/* Initiate send/recv to/from others. */
total_reqs = (((max_outstanding_reqs > (size - 1)) ||
(max_outstanding_reqs <= 0)) ?
(size - 1) : (max_outstanding_reqs));
reqs = coll_base_comm_get_reqs(module->base_data, 2 * total_reqs);
if (NULL == reqs) { error = -1; line = __LINE__; goto error_hndl; }
prcv = (char *) rbuf;
psnd = (char *) sbuf;
/* Post first batch or ireceive and isend requests */
for (nreqs = 0, nrreqs = 0, ri = (rank + 1) % size; nreqs < total_reqs;
ri = (ri + 1) % size, ++nreqs, ++nrreqs) {
error =
MCA_PML_CALL(irecv
(prcv + (ptrdiff_t)ri * rext, rcount, rdtype, ri,
MCA_COLL_BASE_TAG_ALLTOALL, comm, &reqs[nreqs]));
if (MPI_SUCCESS != error) { line = __LINE__; goto error_hndl; }
}
for ( nsreqs = 0, si = (rank + size - 1) % size; nreqs < 2 * total_reqs;
si = (si + size - 1) % size, ++nreqs, ++nsreqs) {
error =
MCA_PML_CALL(isend
(psnd + (ptrdiff_t)si * sext, scount, sdtype, si,
MCA_COLL_BASE_TAG_ALLTOALL,
MCA_PML_BASE_SEND_STANDARD, comm, &reqs[nreqs]));
if (MPI_SUCCESS != error) { line = __LINE__; goto error_hndl; }
}
/* Wait for requests to complete */
if (nreqs == 2 * (size - 1)) {
/* Optimization for the case when all requests have been posted */
error = ompi_request_wait_all(nreqs, reqs, MPI_STATUSES_IGNORE);
if (MPI_SUCCESS != error) { line = __LINE__; goto error_hndl; }
} else {
/* As requests complete, replace them with corresponding requests:
- wait for any request to complete, mark the request as
MPI_REQUEST_NULL
- If it was a receive request, replace it with new irecv request
(if any)
- if it was a send request, replace it with new isend request (if any)
*/
int ncreqs = 0;
while (ncreqs < 2 * (size - 1)) {
int completed;
error = ompi_request_wait_any(2 * total_reqs, reqs, &completed,
MPI_STATUS_IGNORE);
if (MPI_SUCCESS != error) { line = __LINE__; goto error_hndl; }
reqs[completed] = MPI_REQUEST_NULL;
ncreqs++;
if (completed < total_reqs) {
if (nrreqs < (size - 1)) {
error =
MCA_PML_CALL(irecv
(prcv + (ptrdiff_t)ri * rext, rcount, rdtype, ri,
MCA_COLL_BASE_TAG_ALLTOALL, comm,
&reqs[completed]));
if (MPI_SUCCESS != error) { line = __LINE__; goto error_hndl; }
++nrreqs;
ri = (ri + 1) % size;
}
} else {
if (nsreqs < (size - 1)) {
error = MCA_PML_CALL(isend
(psnd + (ptrdiff_t)si * sext, scount, sdtype, si,
MCA_COLL_BASE_TAG_ALLTOALL,
MCA_PML_BASE_SEND_STANDARD, comm,
&reqs[completed]));
++nsreqs;
si = (si + size - 1) % size;
}
}
}
}
/* All done */
return MPI_SUCCESS;
error_hndl:
OPAL_OUTPUT((ompi_coll_base_framework.framework_output,
"%s:%4d\tError occurred %d, rank %2d", __FILE__, line, error,
rank));
ompi_coll_base_free_reqs(reqs, 2 * total_reqs);
return error;
}
int
mca_fcoll_dynamic_file_write_all (ompio_file_t *fh,
const void *buf,
int count,
struct ompi_datatype_t *datatype,
ompi_status_public_t *status)
{
MPI_Aint total_bytes_written = 0; /* total bytes that have been written*/
MPI_Aint total_bytes = 0; /* total bytes to be written */
MPI_Aint bytes_to_write_in_cycle = 0; /* left to be written in a cycle*/
MPI_Aint bytes_per_cycle = 0; /* total written in each cycle by each process*/
int index = 0;
int cycles = 0;
int i=0, j=0, l=0;
int n=0; /* current position in total_bytes_per_process array */
MPI_Aint bytes_remaining = 0; /* how many bytes have been written from the current
value from total_bytes_per_process */
int bytes_sent = 0, ret =0;
int blocks=0, entries_per_aggregator=0;
/* iovec structure and count of the buffer passed in */
uint32_t iov_count = 0;
struct iovec *decoded_iov = NULL;
int iov_index = 0;
char *send_buf = NULL;
size_t current_position = 0;
struct iovec *local_iov_array=NULL, *global_iov_array=NULL;
mca_io_ompio_local_io_array *file_offsets_for_agg=NULL;
/* global iovec at the writers that contain the iovecs created from
file_set_view */
uint32_t total_fview_count = 0;
int local_count = 0, temp_pindex;
int *fview_count = NULL, *disp_index=NULL, *temp_disp_index=NULL;
int current_index = 0, temp_index=0;
char *global_buf = NULL;
MPI_Aint global_count = 0;
/* array that contains the sorted indices of the global_iov */
int *sorted = NULL, *sorted_file_offsets=NULL;
int *displs = NULL;
int dynamic_num_io_procs;
size_t max_data = 0, datatype_size = 0;
int **blocklen_per_process=NULL;
MPI_Aint **displs_per_process=NULL, *memory_displacements=NULL;
ompi_datatype_t **recvtype = NULL;
MPI_Aint *total_bytes_per_process = NULL;
MPI_Request send_req=NULL, *recv_req=NULL;
int my_aggregator=-1;
bool sendbuf_is_contiguous = false;
size_t ftype_size;
ptrdiff_t ftype_extent, lb;
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
double write_time = 0.0, start_write_time = 0.0, end_write_time = 0.0;
double comm_time = 0.0, start_comm_time = 0.0, end_comm_time = 0.0;
double exch_write = 0.0, start_exch = 0.0, end_exch = 0.0;
mca_common_ompio_print_entry nentry;
#endif
opal_datatype_type_size ( &datatype->super, &ftype_size );
opal_datatype_get_extent ( &datatype->super, &lb, &ftype_extent );
/**************************************************************************
** 1. In case the data is not contigous in memory, decode it into an iovec
**************************************************************************/
if ( ( ftype_extent == (ptrdiff_t) ftype_size) &&
opal_datatype_is_contiguous_memory_layout(&datatype->super,1) &&
0 == lb ) {
sendbuf_is_contiguous = true;
}
if (! sendbuf_is_contiguous ) {
ret = mca_common_ompio_decode_datatype ((struct ompio_file_t *) fh,
datatype,
count,
buf,
&max_data,
&decoded_iov,
&iov_count);
if (OMPI_SUCCESS != ret ){
goto exit;
}
}
else {
max_data = count * datatype->super.size;
}
if ( MPI_STATUS_IGNORE != status ) {
status->_ucount = max_data;
}
dynamic_num_io_procs = fh->f_get_mca_parameter_value ( "num_aggregators", strlen ("num_aggregators"));
if ( OMPI_ERR_MAX == dynamic_num_io_procs ) {
ret = OMPI_ERROR;
goto exit;
}
ret = mca_common_ompio_set_aggregator_props ((struct ompio_file_t *) fh,
dynamic_num_io_procs,
max_data);
if (OMPI_SUCCESS != ret){
goto exit;
}
my_aggregator = fh->f_procs_in_group[0];
/**************************************************************************
** 2. Determine the total amount of data to be written
**************************************************************************/
total_bytes_per_process = (MPI_Aint*)malloc
(fh->f_procs_per_group*sizeof(MPI_Aint));
if (NULL == total_bytes_per_process) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_comm_time = MPI_Wtime();
#endif
ret = ompi_fcoll_base_coll_allgather_array (&max_data,
1,
MPI_LONG,
total_bytes_per_process,
1,
MPI_LONG,
0,
fh->f_procs_in_group,
fh->f_procs_per_group,
fh->f_comm);
if( OMPI_SUCCESS != ret){
goto exit;
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_comm_time = MPI_Wtime();
comm_time += (end_comm_time - start_comm_time);
#endif
for (i=0 ; i<fh->f_procs_per_group ; i++) {
total_bytes += total_bytes_per_process[i];
}
if (NULL != total_bytes_per_process) {
free (total_bytes_per_process);
total_bytes_per_process = NULL;
}
/*********************************************************************
*** 3. Generate the local offsets/lengths array corresponding to
*** this write operation
********************************************************************/
ret = fh->f_generate_current_file_view( (struct ompio_file_t *) fh,
max_data,
&local_iov_array,
&local_count);
if (ret != OMPI_SUCCESS){
goto exit;
}
#if DEBUG_ON
for (i=0 ; i<local_count ; i++) {
printf("%d: OFFSET: %d LENGTH: %ld\n",
fh->f_rank,
local_iov_array[i].iov_base,
local_iov_array[i].iov_len);
}
#endif
/*************************************************************
*** 4. Allgather the offset/lengths array from all processes
*************************************************************/
fview_count = (int *) malloc (fh->f_procs_per_group * sizeof (int));
if (NULL == fview_count) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_comm_time = MPI_Wtime();
#endif
ret = ompi_fcoll_base_coll_allgather_array (&local_count,
1,
MPI_INT,
fview_count,
1,
MPI_INT,
0,
fh->f_procs_in_group,
fh->f_procs_per_group,
fh->f_comm);
if( OMPI_SUCCESS != ret){
goto exit;
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_comm_time = MPI_Wtime();
comm_time += (end_comm_time - start_comm_time);
#endif
displs = (int*) malloc (fh->f_procs_per_group * sizeof (int));
if (NULL == displs) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
displs[0] = 0;
total_fview_count = fview_count[0];
for (i=1 ; i<fh->f_procs_per_group ; i++) {
total_fview_count += fview_count[i];
displs[i] = displs[i-1] + fview_count[i-1];
}
#if DEBUG_ON
printf("total_fview_count : %d\n", total_fview_count);
if (my_aggregator == fh->f_rank) {
for (i=0 ; i<fh->f_procs_per_group ; i++) {
printf ("%d: PROCESS: %d ELEMENTS: %d DISPLS: %d\n",
fh->f_rank,
i,
fview_count[i],
displs[i]);
}
}
#endif
/* allocate the global iovec */
if (0 != total_fview_count) {
global_iov_array = (struct iovec*) malloc (total_fview_count *
sizeof(struct iovec));
if (NULL == global_iov_array){
opal_output(1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_comm_time = MPI_Wtime();
#endif
ret = ompi_fcoll_base_coll_allgatherv_array (local_iov_array,
local_count,
fh->f_iov_type,
global_iov_array,
fview_count,
displs,
fh->f_iov_type,
0,
fh->f_procs_in_group,
fh->f_procs_per_group,
fh->f_comm);
if (OMPI_SUCCESS != ret){
goto exit;
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_comm_time = MPI_Wtime();
comm_time += (end_comm_time - start_comm_time);
#endif
/****************************************************************************************
*** 5. Sort the global offset/lengths list based on the offsets.
*** The result of the sort operation is the 'sorted', an integer array,
*** which contains the indexes of the global_iov_array based on the offset.
*** For example, if global_iov_array[x].offset is followed by global_iov_array[y].offset
*** in the file, and that one is followed by global_iov_array[z].offset, than
*** sorted[0] = x, sorted[1]=y and sorted[2]=z;
******************************************************************************************/
if (0 != total_fview_count) {
sorted = (int *)malloc (total_fview_count * sizeof(int));
if (NULL == sorted) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
ompi_fcoll_base_sort_iovec (global_iov_array, total_fview_count, sorted);
}
if (NULL != local_iov_array){
free(local_iov_array);
local_iov_array = NULL;
}
if (NULL != displs){
free(displs);
displs=NULL;
}
#if DEBUG_ON
if (my_aggregator == fh->f_rank) {
uint32_t tv=0;
for (tv=0 ; tv<total_fview_count ; tv++) {
printf("%d: OFFSET: %lld LENGTH: %ld\n",
fh->f_rank,
global_iov_array[sorted[tv]].iov_base,
global_iov_array[sorted[tv]].iov_len);
}
}
#endif
/*************************************************************
*** 6. Determine the number of cycles required to execute this
*** operation
*************************************************************/
bytes_per_cycle = fh->f_bytes_per_agg;
cycles = ceil((double)total_bytes/bytes_per_cycle);
if (my_aggregator == fh->f_rank) {
disp_index = (int *)malloc (fh->f_procs_per_group * sizeof (int));
if (NULL == disp_index) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
blocklen_per_process = (int **)calloc (fh->f_procs_per_group, sizeof (int*));
if (NULL == blocklen_per_process) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
displs_per_process = (MPI_Aint **)calloc (fh->f_procs_per_group, sizeof (MPI_Aint*));
if (NULL == displs_per_process) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
recv_req = (MPI_Request *)malloc ((fh->f_procs_per_group)*sizeof(MPI_Request));
if ( NULL == recv_req ) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
global_buf = (char *) malloc (bytes_per_cycle);
if (NULL == global_buf){
opal_output(1, "OUT OF MEMORY");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
recvtype = (ompi_datatype_t **) malloc (fh->f_procs_per_group * sizeof(ompi_datatype_t *));
if (NULL == recvtype) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
for(l=0;l<fh->f_procs_per_group;l++){
recvtype[l] = MPI_DATATYPE_NULL;
}
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_exch = MPI_Wtime();
#endif
n = 0;
bytes_remaining = 0;
current_index = 0;
for (index = 0; index < cycles; index++) {
/**********************************************************************
*** 7a. Getting ready for next cycle: initializing and freeing buffers
**********************************************************************/
if (my_aggregator == fh->f_rank) {
if (NULL != fh->f_io_array) {
free (fh->f_io_array);
fh->f_io_array = NULL;
}
fh->f_num_of_io_entries = 0;
if (NULL != recvtype){
for (i =0; i< fh->f_procs_per_group; i++) {
if ( MPI_DATATYPE_NULL != recvtype[i] ) {
ompi_datatype_destroy(&recvtype[i]);
recvtype[i] = MPI_DATATYPE_NULL;
}
}
}
for(l=0;l<fh->f_procs_per_group;l++){
disp_index[l] = 1;
free(blocklen_per_process[l]);
free(displs_per_process[l]);
blocklen_per_process[l] = (int *) calloc (1, sizeof(int));
displs_per_process[l] = (MPI_Aint *) calloc (1, sizeof(MPI_Aint));
if (NULL == displs_per_process[l] || NULL == blocklen_per_process[l]){
opal_output (1, "OUT OF MEMORY for displs\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
}
if (NULL != sorted_file_offsets){
free(sorted_file_offsets);
sorted_file_offsets = NULL;
}
if(NULL != file_offsets_for_agg){
free(file_offsets_for_agg);
file_offsets_for_agg = NULL;
}
if (NULL != memory_displacements){
free(memory_displacements);
memory_displacements = NULL;
}
} /* (my_aggregator == fh->f_rank */
/**************************************************************************
*** 7b. Determine the number of bytes to be actually written in this cycle
**************************************************************************/
if (cycles-1 == index) {
bytes_to_write_in_cycle = total_bytes - bytes_per_cycle*index;
}
else {
bytes_to_write_in_cycle = bytes_per_cycle;
}
#if DEBUG_ON
if (my_aggregator == fh->f_rank) {
printf ("****%d: CYCLE %d Bytes %lld**********\n",
fh->f_rank,
index,
bytes_to_write_in_cycle);
}
#endif
/**********************************************************
**Gather the Data from all the processes at the writers **
*********************************************************/
#if DEBUG_ON
printf("bytes_to_write_in_cycle: %ld, cycle : %d\n", bytes_to_write_in_cycle,
index);
#endif
/*****************************************************************
*** 7c. Calculate how much data will be contributed in this cycle
*** by each process
*****************************************************************/
bytes_sent = 0;
/* The blocklen and displs calculation only done at aggregators!*/
while (bytes_to_write_in_cycle) {
/* This next block identifies which process is the holder
** of the sorted[current_index] element;
*/
blocks = fview_count[0];
for (j=0 ; j<fh->f_procs_per_group ; j++) {
if (sorted[current_index] < blocks) {
n = j;
break;
}
else {
blocks += fview_count[j+1];
}
}
if (bytes_remaining) {
/* Finish up a partially used buffer from the previous cycle */
if (bytes_remaining <= bytes_to_write_in_cycle) {
/* The data fits completely into the block */
if (my_aggregator == fh->f_rank) {
blocklen_per_process[n][disp_index[n] - 1] = bytes_remaining;
displs_per_process[n][disp_index[n] - 1] =
(ptrdiff_t)global_iov_array[sorted[current_index]].iov_base +
(global_iov_array[sorted[current_index]].iov_len
- bytes_remaining);
/* In this cases the length is consumed so allocating for
next displacement and blocklength*/
blocklen_per_process[n] = (int *) realloc
((void *)blocklen_per_process[n], (disp_index[n]+1)*sizeof(int));
displs_per_process[n] = (MPI_Aint *) realloc
((void *)displs_per_process[n], (disp_index[n]+1)*sizeof(MPI_Aint));
blocklen_per_process[n][disp_index[n]] = 0;
displs_per_process[n][disp_index[n]] = 0;
disp_index[n] += 1;
}
if (fh->f_procs_in_group[n] == fh->f_rank) {
bytes_sent += bytes_remaining;
}
current_index ++;
bytes_to_write_in_cycle -= bytes_remaining;
bytes_remaining = 0;
continue;
}
else {
/* the remaining data from the previous cycle is larger than the
bytes_to_write_in_cycle, so we have to segment again */
if (my_aggregator == fh->f_rank) {
blocklen_per_process[n][disp_index[n] - 1] = bytes_to_write_in_cycle;
displs_per_process[n][disp_index[n] - 1] =
(ptrdiff_t)global_iov_array[sorted[current_index]].iov_base +
(global_iov_array[sorted[current_index]].iov_len
- bytes_remaining);
}
if (fh->f_procs_in_group[n] == fh->f_rank) {
bytes_sent += bytes_to_write_in_cycle;
}
bytes_remaining -= bytes_to_write_in_cycle;
bytes_to_write_in_cycle = 0;
break;
}
}
else {
/* No partially used entry available, have to start a new one */
if (bytes_to_write_in_cycle <
(MPI_Aint) global_iov_array[sorted[current_index]].iov_len) {
/* This entry has more data than we can sendin one cycle */
if (my_aggregator == fh->f_rank) {
blocklen_per_process[n][disp_index[n] - 1] = bytes_to_write_in_cycle;
displs_per_process[n][disp_index[n] - 1] =
(ptrdiff_t)global_iov_array[sorted[current_index]].iov_base ;
}
if (fh->f_procs_in_group[n] == fh->f_rank) {
bytes_sent += bytes_to_write_in_cycle;
}
bytes_remaining = global_iov_array[sorted[current_index]].iov_len -
bytes_to_write_in_cycle;
bytes_to_write_in_cycle = 0;
break;
}
else {
/* Next data entry is less than bytes_to_write_in_cycle */
if (my_aggregator == fh->f_rank) {
blocklen_per_process[n][disp_index[n] - 1] =
global_iov_array[sorted[current_index]].iov_len;
displs_per_process[n][disp_index[n] - 1] = (ptrdiff_t)
global_iov_array[sorted[current_index]].iov_base;
/*realloc for next blocklength
and assign this displacement and check for next displs as
the total length of this entry has been consumed!*/
blocklen_per_process[n] =
(int *) realloc ((void *)blocklen_per_process[n], (disp_index[n]+1)*sizeof(int));
displs_per_process[n] = (MPI_Aint *)realloc
((void *)displs_per_process[n], (disp_index[n]+1)*sizeof(MPI_Aint));
blocklen_per_process[n][disp_index[n]] = 0;
displs_per_process[n][disp_index[n]] = 0;
disp_index[n] += 1;
}
if (fh->f_procs_in_group[n] == fh->f_rank) {
bytes_sent += global_iov_array[sorted[current_index]].iov_len;
}
bytes_to_write_in_cycle -=
global_iov_array[sorted[current_index]].iov_len;
current_index ++;
continue;
}
}
}
/*************************************************************************
*** 7d. Calculate the displacement on where to put the data and allocate
*** the recieve buffer (global_buf)
*************************************************************************/
if (my_aggregator == fh->f_rank) {
entries_per_aggregator=0;
for (i=0;i<fh->f_procs_per_group; i++){
for (j=0;j<disp_index[i];j++){
if (blocklen_per_process[i][j] > 0)
entries_per_aggregator++ ;
}
}
#if DEBUG_ON
printf("%d: cycle: %d, bytes_sent: %d\n ",fh->f_rank,index,
bytes_sent);
printf("%d : Entries per aggregator : %d\n",fh->f_rank,entries_per_aggregator);
#endif
if (entries_per_aggregator > 0){
file_offsets_for_agg = (mca_io_ompio_local_io_array *)
malloc(entries_per_aggregator*sizeof(mca_io_ompio_local_io_array));
if (NULL == file_offsets_for_agg) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
sorted_file_offsets = (int *)
malloc (entries_per_aggregator*sizeof(int));
if (NULL == sorted_file_offsets){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
/*Moving file offsets to an IO array!*/
temp_index = 0;
for (i=0;i<fh->f_procs_per_group; i++){
for(j=0;j<disp_index[i];j++){
if (blocklen_per_process[i][j] > 0){
file_offsets_for_agg[temp_index].length =
blocklen_per_process[i][j];
file_offsets_for_agg[temp_index].process_id = i;
file_offsets_for_agg[temp_index].offset =
displs_per_process[i][j];
temp_index++;
#if DEBUG_ON
printf("************Cycle: %d, Aggregator: %d ***************\n",
index+1,fh->f_rank);
printf("%d sends blocklen[%d]: %d, disp[%d]: %ld to %d\n",
fh->f_procs_in_group[i],j,
blocklen_per_process[i][j],j,
displs_per_process[i][j],
fh->f_rank);
#endif
}
}
}
}
else{
continue;
}
/* Sort the displacements for each aggregator*/
local_heap_sort (file_offsets_for_agg,
entries_per_aggregator,
sorted_file_offsets);
/*create contiguous memory displacements
based on blocklens on the same displs array
and map it to this aggregator's actual
file-displacements (this is in the io-array created above)*/
memory_displacements = (MPI_Aint *) malloc
(entries_per_aggregator * sizeof(MPI_Aint));
memory_displacements[sorted_file_offsets[0]] = 0;
for (i=1; i<entries_per_aggregator; i++){
memory_displacements[sorted_file_offsets[i]] =
memory_displacements[sorted_file_offsets[i-1]] +
file_offsets_for_agg[sorted_file_offsets[i-1]].length;
}
temp_disp_index = (int *)calloc (1, fh->f_procs_per_group * sizeof (int));
if (NULL == temp_disp_index) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
/*Now update the displacements array with memory offsets*/
global_count = 0;
for (i=0;i<entries_per_aggregator;i++){
temp_pindex =
file_offsets_for_agg[sorted_file_offsets[i]].process_id;
displs_per_process[temp_pindex][temp_disp_index[temp_pindex]] =
memory_displacements[sorted_file_offsets[i]];
if (temp_disp_index[temp_pindex] < disp_index[temp_pindex])
temp_disp_index[temp_pindex] += 1;
else{
printf("temp_disp_index[%d]: %d is greater than disp_index[%d]: %d\n",
temp_pindex, temp_disp_index[temp_pindex],
temp_pindex, disp_index[temp_pindex]);
}
global_count +=
file_offsets_for_agg[sorted_file_offsets[i]].length;
}
if (NULL != temp_disp_index){
free(temp_disp_index);
temp_disp_index = NULL;
}
#if DEBUG_ON
printf("************Cycle: %d, Aggregator: %d ***************\n",
index+1,fh->f_rank);
for (i=0;i<fh->f_procs_per_group; i++){
for(j=0;j<disp_index[i];j++){
if (blocklen_per_process[i][j] > 0){
printf("%d sends blocklen[%d]: %d, disp[%d]: %ld to %d\n",
fh->f_procs_in_group[i],j,
blocklen_per_process[i][j],j,
displs_per_process[i][j],
fh->f_rank);
}
}
}
printf("************Cycle: %d, Aggregator: %d ***************\n",
index+1,fh->f_rank);
for (i=0; i<entries_per_aggregator;i++){
printf("%d: OFFSET: %lld LENGTH: %ld, Mem-offset: %ld\n",
file_offsets_for_agg[sorted_file_offsets[i]].process_id,
file_offsets_for_agg[sorted_file_offsets[i]].offset,
file_offsets_for_agg[sorted_file_offsets[i]].length,
memory_displacements[sorted_file_offsets[i]]);
}
printf("%d : global_count : %ld, bytes_sent : %d\n",
fh->f_rank,global_count, bytes_sent);
#endif
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_comm_time = MPI_Wtime();
#endif
/*************************************************************************
*** 7e. Perform the actual communication
*************************************************************************/
for (i=0;i<fh->f_procs_per_group; i++) {
recv_req[i] = MPI_REQUEST_NULL;
if ( 0 < disp_index[i] ) {
ompi_datatype_create_hindexed(disp_index[i],
blocklen_per_process[i],
displs_per_process[i],
MPI_BYTE,
&recvtype[i]);
ompi_datatype_commit(&recvtype[i]);
opal_datatype_type_size(&recvtype[i]->super, &datatype_size);
if (datatype_size){
ret = MCA_PML_CALL(irecv(global_buf,
1,
recvtype[i],
fh->f_procs_in_group[i],
123,
fh->f_comm,
&recv_req[i]));
if (OMPI_SUCCESS != ret){
goto exit;
}
}
}
}
} /* end if (my_aggregator == fh->f_rank ) */
if ( sendbuf_is_contiguous ) {
send_buf = &((char*)buf)[total_bytes_written];
}
else if (bytes_sent) {
/* allocate a send buffer and copy the data that needs
to be sent into it in case the data is non-contigous
in memory */
ptrdiff_t mem_address;
size_t remaining = 0;
size_t temp_position = 0;
send_buf = malloc (bytes_sent);
if (NULL == send_buf) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
remaining = bytes_sent;
while (remaining) {
mem_address = (ptrdiff_t)
(decoded_iov[iov_index].iov_base) + current_position;
if (remaining >=
(decoded_iov[iov_index].iov_len - current_position)) {
memcpy (send_buf+temp_position,
(IOVBASE_TYPE *)mem_address,
decoded_iov[iov_index].iov_len - current_position);
remaining = remaining -
(decoded_iov[iov_index].iov_len - current_position);
temp_position = temp_position +
(decoded_iov[iov_index].iov_len - current_position);
iov_index = iov_index + 1;
current_position = 0;
}
else {
memcpy (send_buf+temp_position,
(IOVBASE_TYPE *) mem_address,
remaining);
current_position = current_position + remaining;
remaining = 0;
}
}
}
total_bytes_written += bytes_sent;
/* Gather the sendbuf from each process in appropritate locations in
aggregators*/
if (bytes_sent){
ret = MCA_PML_CALL(isend(send_buf,
bytes_sent,
MPI_BYTE,
my_aggregator,
123,
MCA_PML_BASE_SEND_STANDARD,
fh->f_comm,
&send_req));
if ( OMPI_SUCCESS != ret ){
goto exit;
}
ret = ompi_request_wait(&send_req, MPI_STATUS_IGNORE);
if (OMPI_SUCCESS != ret){
goto exit;
}
}
if (my_aggregator == fh->f_rank) {
ret = ompi_request_wait_all (fh->f_procs_per_group,
recv_req,
MPI_STATUS_IGNORE);
if (OMPI_SUCCESS != ret){
goto exit;
}
}
#if DEBUG_ON
if (my_aggregator == fh->f_rank){
printf("************Cycle: %d, Aggregator: %d ***************\n",
index+1,fh->f_rank);
for (i=0 ; i<global_count/4 ; i++)
printf (" RECV %d \n",((int *)global_buf)[i]);
}
#endif
if (! sendbuf_is_contiguous) {
if (NULL != send_buf) {
free (send_buf);
send_buf = NULL;
}
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_comm_time = MPI_Wtime();
comm_time += (end_comm_time - start_comm_time);
#endif
/**********************************************************
*** 7f. Create the io array, and pass it to fbtl
*********************************************************/
if (my_aggregator == fh->f_rank) {
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_write_time = MPI_Wtime();
#endif
fh->f_io_array = (mca_common_ompio_io_array_t *) malloc
(entries_per_aggregator * sizeof (mca_common_ompio_io_array_t));
if (NULL == fh->f_io_array) {
opal_output(1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
fh->f_num_of_io_entries = 0;
/*First entry for every aggregator*/
fh->f_io_array[0].offset =
(IOVBASE_TYPE *)(intptr_t)file_offsets_for_agg[sorted_file_offsets[0]].offset;
fh->f_io_array[0].length =
file_offsets_for_agg[sorted_file_offsets[0]].length;
fh->f_io_array[0].memory_address =
global_buf+memory_displacements[sorted_file_offsets[0]];
fh->f_num_of_io_entries++;
for (i=1;i<entries_per_aggregator;i++){
/* If the enrties are contiguous merge them,
else make a new entry */
if (file_offsets_for_agg[sorted_file_offsets[i-1]].offset +
file_offsets_for_agg[sorted_file_offsets[i-1]].length ==
file_offsets_for_agg[sorted_file_offsets[i]].offset){
fh->f_io_array[fh->f_num_of_io_entries - 1].length +=
file_offsets_for_agg[sorted_file_offsets[i]].length;
}
else {
fh->f_io_array[fh->f_num_of_io_entries].offset =
(IOVBASE_TYPE *)(intptr_t)file_offsets_for_agg[sorted_file_offsets[i]].offset;
fh->f_io_array[fh->f_num_of_io_entries].length =
file_offsets_for_agg[sorted_file_offsets[i]].length;
fh->f_io_array[fh->f_num_of_io_entries].memory_address =
global_buf+memory_displacements[sorted_file_offsets[i]];
fh->f_num_of_io_entries++;
}
}
#if DEBUG_ON
printf("*************************** %d\n", fh->f_num_of_io_entries);
for (i=0 ; i<fh->f_num_of_io_entries ; i++) {
printf(" ADDRESS: %p OFFSET: %ld LENGTH: %ld\n",
fh->f_io_array[i].memory_address,
(ptrdiff_t)fh->f_io_array[i].offset,
fh->f_io_array[i].length);
}
#endif
if (fh->f_num_of_io_entries) {
if ( 0 > fh->f_fbtl->fbtl_pwritev (fh)) {
opal_output (1, "WRITE FAILED\n");
ret = OMPI_ERROR;
goto exit;
}
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_write_time = MPI_Wtime();
write_time += end_write_time - start_write_time;
#endif
} /* end if (my_aggregator == fh->f_rank) */
} /* end for (index = 0; index < cycles; index++) */
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_exch = MPI_Wtime();
exch_write += end_exch - start_exch;
nentry.time[0] = write_time;
nentry.time[1] = comm_time;
nentry.time[2] = exch_write;
if (my_aggregator == fh->f_rank)
nentry.aggregator = 1;
else
nentry.aggregator = 0;
nentry.nprocs_for_coll = dynamic_num_io_procs;
if (!mca_common_ompio_full_print_queue(fh->f_coll_write_time)){
mca_common_ompio_register_print_entry(fh->f_coll_write_time,
nentry);
}
#endif
exit :
if (my_aggregator == fh->f_rank) {
if (NULL != sorted_file_offsets){
free(sorted_file_offsets);
sorted_file_offsets = NULL;
}
if(NULL != file_offsets_for_agg){
free(file_offsets_for_agg);
file_offsets_for_agg = NULL;
}
if (NULL != memory_displacements){
free(memory_displacements);
memory_displacements = NULL;
}
if (NULL != recvtype){
for (i =0; i< fh->f_procs_per_group; i++) {
if ( MPI_DATATYPE_NULL != recvtype[i] ) {
ompi_datatype_destroy(&recvtype[i]);
}
}
free(recvtype);
recvtype=NULL;
}
if (NULL != fh->f_io_array) {
free (fh->f_io_array);
fh->f_io_array = NULL;
}
if (NULL != disp_index){
free(disp_index);
disp_index = NULL;
}
if (NULL != recvtype){
free(recvtype);
recvtype=NULL;
}
if (NULL != recv_req){
free(recv_req);
recv_req = NULL;
}
if (NULL != global_buf) {
free (global_buf);
global_buf = NULL;
}
for(l=0;l<fh->f_procs_per_group;l++){
if (NULL != blocklen_per_process){
free(blocklen_per_process[l]);
}
if (NULL != displs_per_process){
free(displs_per_process[l]);
}
}
free(blocklen_per_process);
free(displs_per_process);
}
if (NULL != displs){
free(displs);
displs=NULL;
}
if (! sendbuf_is_contiguous) {
if (NULL != send_buf) {
free (send_buf);
send_buf = NULL;
}
}
if (NULL != global_buf) {
free (global_buf);
global_buf = NULL;
}
if (NULL != sorted) {
free (sorted);
sorted = NULL;
}
if (NULL != global_iov_array) {
free (global_iov_array);
global_iov_array = NULL;
}
if (NULL != fview_count) {
free (fview_count);
fview_count = NULL;
}
if (NULL != decoded_iov) {
free (decoded_iov);
decoded_iov = NULL;
}
return OMPI_SUCCESS;
}
/* Arguments not used in this implementation:
* - send_first
*/
static int ompi_comm_allreduce_intra_bridge (int *inbuf, int *outbuf,
int count, struct ompi_op_t *op,
ompi_communicator_t *comm,
ompi_communicator_t *bcomm,
void* lleader, void* rleader,
int send_first )
{
int *tmpbuf=NULL;
int local_rank;
int i;
int rc;
int local_leader, remote_leader;
local_leader = (*((int*)lleader));
remote_leader = (*((int*)rleader));
if ( &ompi_mpi_op_sum.op != op && &ompi_mpi_op_prod.op != op &&
&ompi_mpi_op_max.op != op && &ompi_mpi_op_min.op != op ) {
return MPI_ERR_OP;
}
local_rank = ompi_comm_rank ( comm );
tmpbuf = (int *) malloc ( count * sizeof(int));
if ( NULL == tmpbuf ) {
rc = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
/* Intercomm_create */
rc = comm->c_coll.coll_allreduce ( inbuf, tmpbuf, count, MPI_INT,
op, comm, comm->c_coll.coll_allreduce_module );
if ( OMPI_SUCCESS != rc ) {
goto exit;
}
if (local_rank == local_leader ) {
MPI_Request req;
rc = MCA_PML_CALL(irecv ( outbuf, count, MPI_INT, remote_leader,
OMPI_COMM_ALLREDUCE_TAG,
bcomm, &req));
if ( OMPI_SUCCESS != rc ) {
goto exit;
}
rc = MCA_PML_CALL(send (tmpbuf, count, MPI_INT, remote_leader,
OMPI_COMM_ALLREDUCE_TAG,
MCA_PML_BASE_SEND_STANDARD, bcomm));
if ( OMPI_SUCCESS != rc ) {
goto exit;
}
rc = ompi_request_wait_all ( 1, &req, MPI_STATUS_IGNORE);
if ( OMPI_SUCCESS != rc ) {
goto exit;
}
if ( &ompi_mpi_op_max.op == op ) {
for ( i = 0 ; i < count; i++ ) {
if (tmpbuf[i] > outbuf[i]) {
outbuf[i] = tmpbuf[i];
}
}
}
else if ( &ompi_mpi_op_min.op == op ) {
for ( i = 0 ; i < count; i++ ) {
if (tmpbuf[i] < outbuf[i]) {
outbuf[i] = tmpbuf[i];
}
}
}
else if ( &ompi_mpi_op_sum.op == op ) {
for ( i = 0 ; i < count; i++ ) {
outbuf[i] += tmpbuf[i];
}
}
else if ( &ompi_mpi_op_prod.op == op ) {
for ( i = 0 ; i < count; i++ ) {
outbuf[i] *= tmpbuf[i];
}
}
}
rc = comm->c_coll.coll_bcast ( outbuf, count, MPI_INT, local_leader,
comm, comm->c_coll.coll_bcast_module );
exit:
if (NULL != tmpbuf ) {
free (tmpbuf);
}
return (rc);
}
/*
* allgatherv_inter
*
* Function: - allgatherv using other MPI collectives
* Accepts: - same as MPI_Allgatherv()
* Returns: - MPI_SUCCESS or error code
*/
int
mca_coll_inter_allgatherv_inter(void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
void *rbuf, int *rcounts, int *disps,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int i, rank, size, size_local, total=0, err;
int *count=NULL,*displace=NULL;
char *ptmp=NULL;
MPI_Aint incr;
MPI_Aint extent;
MPI_Aint lb;
ompi_datatype_t *ndtype;
ompi_request_t *req[2];
rank = ompi_comm_rank(comm);
size_local = ompi_comm_size(comm->c_local_comm);
size = ompi_comm_remote_size(comm);
if (0 == rank) {
count = (int *)malloc(sizeof(int) * size_local);
if (NULL == count) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
displace = (int *)malloc(sizeof(int) * size_local);
if (NULL == displace) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
}
/* Local gather to get the scount of each process */
err = comm->c_local_comm->c_coll.coll_gather(&scount, 1, MPI_INT,
count, 1, MPI_INT,
0, comm->c_local_comm,
comm->c_local_comm->c_coll.coll_gather_module);
if (OMPI_SUCCESS != err) {
return err;
}
if(0 == rank) {
displace[0] = 0;
for (i = 1; i < size_local; i++) {
displace[i] = displace[i-1] + count[i-1];
}
/* Perform the gatherv locally with the first process as root */
err = ompi_ddt_get_extent(sdtype, &lb, &extent);
if (OMPI_SUCCESS != err) {
return OMPI_ERROR;
}
incr = 0;
for (i = 0; i < size_local; i++) {
incr = incr + extent*count[i];
}
ptmp = (char*)malloc(incr);
if (NULL == ptmp) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
}
err = comm->c_local_comm->c_coll.coll_gatherv(sbuf, scount, sdtype,
ptmp, count, displace,
sdtype,0, comm->c_local_comm,
comm->c_local_comm->c_coll.coll_gatherv_module);
if (OMPI_SUCCESS != err) {
return err;
}
ompi_ddt_create_indexed(size,rcounts,disps,rdtype,&ndtype);
ompi_ddt_commit(&ndtype);
if (0 == rank) {
for (i = 0; i < size_local; i++) {
total = total + count[i];
}
/* Exchange data between roots */
err = MCA_PML_CALL(irecv(rbuf, 1, ndtype, 0,
MCA_COLL_BASE_TAG_ALLGATHERV, comm,
&(req[0])));
if (OMPI_SUCCESS != err) {
return err;
}
err = MCA_PML_CALL(isend(ptmp, total, sdtype, 0,
MCA_COLL_BASE_TAG_ALLGATHERV,
MCA_PML_BASE_SEND_STANDARD,
comm, &(req[1])));
if (OMPI_SUCCESS != err) {
return err;
}
err = ompi_request_wait_all(2, req, MPI_STATUSES_IGNORE);
if (OMPI_SUCCESS != err) {
return err;
}
}
/* bcast the message to all the local processes */
err = comm->c_local_comm->c_coll.coll_bcast(rbuf, 1, ndtype,
0, comm->c_local_comm,
comm->c_local_comm->c_coll.coll_bcast_module);
if (OMPI_SUCCESS != err) {
return err;
}
ompi_ddt_destroy(&ndtype);
if (NULL != ptmp) {
free(ptmp);
}
if (NULL != displace) {
free(displace);
}
if (NULL != count) {
free(count);
}
return err;
}
/*
* allgather_inter
*
* Function: - allgather using other MPI collections
* Accepts: - same as MPI_Allgather()
* Returns: - MPI_SUCCESS or error code
*/
int
mca_coll_basic_allgather_inter(void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
void *rbuf, int rcount,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int rank, root = 0, size, rsize, err, i;
char *tmpbuf = NULL, *ptmp;
ptrdiff_t rlb, slb, rextent, sextent, incr;
ompi_request_t *req;
mca_coll_basic_module_t *basic_module = (mca_coll_basic_module_t*) module;
ompi_request_t **reqs = basic_module->mccb_reqs;
rank = ompi_comm_rank(comm);
size = ompi_comm_size(comm);
rsize = ompi_comm_remote_size(comm);
/* Algorithm:
* - a gather to the root in remote group (simultaniously executed,
* thats why we cannot use coll_gather).
* - exchange the temp-results between two roots
* - inter-bcast (again simultanious).
*/
/* Step one: gather operations: */
if (rank != root) {
/* send your data to root */
err = MCA_PML_CALL(send(sbuf, scount, sdtype, root,
MCA_COLL_BASE_TAG_ALLGATHER,
MCA_PML_BASE_SEND_STANDARD, comm));
if (OMPI_SUCCESS != err) {
return err;
}
} else {
/* receive a msg. from all other procs. */
err = ompi_datatype_get_extent(rdtype, &rlb, &rextent);
if (OMPI_SUCCESS != err) {
return err;
}
err = ompi_datatype_get_extent(sdtype, &slb, &sextent);
if (OMPI_SUCCESS != err) {
return err;
}
/* Do a send-recv between the two root procs. to avoid deadlock */
err = MCA_PML_CALL(isend(sbuf, scount, sdtype, 0,
MCA_COLL_BASE_TAG_ALLGATHER,
MCA_PML_BASE_SEND_STANDARD,
comm, &reqs[rsize]));
if (OMPI_SUCCESS != err) {
return err;
}
err = MCA_PML_CALL(irecv(rbuf, rcount, rdtype, 0,
MCA_COLL_BASE_TAG_ALLGATHER, comm,
&reqs[0]));
if (OMPI_SUCCESS != err) {
return err;
}
incr = rextent * rcount;
ptmp = (char *) rbuf + incr;
for (i = 1; i < rsize; ++i, ptmp += incr) {
err = MCA_PML_CALL(irecv(ptmp, rcount, rdtype, i,
MCA_COLL_BASE_TAG_ALLGATHER,
comm, &reqs[i]));
if (MPI_SUCCESS != err) {
return err;
}
}
err = ompi_request_wait_all(rsize + 1, reqs, MPI_STATUSES_IGNORE);
if (OMPI_SUCCESS != err) {
return err;
}
/* Step 2: exchange the resuts between the root processes */
tmpbuf = (char *) malloc(scount * size * sextent);
if (NULL == tmpbuf) {
return err;
}
err = MCA_PML_CALL(isend(rbuf, rsize * rcount, rdtype, 0,
MCA_COLL_BASE_TAG_ALLGATHER,
MCA_PML_BASE_SEND_STANDARD, comm, &req));
if (OMPI_SUCCESS != err) {
goto exit;
}
err = MCA_PML_CALL(recv(tmpbuf, size * scount, sdtype, 0,
MCA_COLL_BASE_TAG_ALLGATHER, comm,
MPI_STATUS_IGNORE));
if (OMPI_SUCCESS != err) {
goto exit;
}
err = ompi_request_wait( &req, MPI_STATUS_IGNORE);
if (OMPI_SUCCESS != err) {
goto exit;
}
}
/* Step 3: bcast the data to the remote group. This
* happens in both groups simultaniously, thus we can
* not use coll_bcast (this would deadlock).
*/
if (rank != root) {
/* post the recv */
err = MCA_PML_CALL(recv(rbuf, rsize * rcount, rdtype, 0,
MCA_COLL_BASE_TAG_ALLGATHER, comm,
MPI_STATUS_IGNORE));
if (OMPI_SUCCESS != err) {
goto exit;
}
} else {
/* Send the data to every other process in the remote group
* except to rank zero. which has it already. */
for (i = 1; i < rsize; i++) {
err = MCA_PML_CALL(isend(tmpbuf, size * scount, sdtype, i,
MCA_COLL_BASE_TAG_ALLGATHER,
MCA_PML_BASE_SEND_STANDARD,
comm, &reqs[i - 1]));
if (OMPI_SUCCESS != err) {
goto exit;
}
}
err = ompi_request_wait_all(rsize - 1, reqs, MPI_STATUSES_IGNORE);
if (OMPI_SUCCESS != err) {
goto exit;
}
}
exit:
if (NULL != tmpbuf) {
free(tmpbuf);
}
return err;
}
static int mca_coll_hierarch_bcast_intra_seg3 (void *buff,
int count,
struct ompi_datatype_t *datatype,
int root,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module,
int segsize )
{
struct ompi_communicator_t *llcomm=NULL;
struct ompi_communicator_t *lcomm=NULL;
mca_coll_hierarch_module_t *hierarch_module = (mca_coll_hierarch_module_t *) module;
int lroot=MPI_UNDEFINED, llroot=MPI_UNDEFINED;
int llrank=MPI_UNDEFINED, llsize=0, rank=0, ret=OMPI_SUCCESS;
int lsize=0, lrank=MPI_UNDEFINED;
MPI_Aint ub=0, typeext=0;
size_t typesize=0;
int i, realsegsize=0, remaining_count=0;
int num_segments=0, segcount=0, segindex=0;
char* tmpbuf = (char *) buff;
ompi_request_t **sreq=NULL, **sreq1=NULL;
ompi_request_t *rreq=MPI_REQUEST_NULL, *rreq1=MPI_REQUEST_NULL;
rank = ompi_comm_rank ( comm );
lcomm = hierarch_module->hier_lcomm;
if ( mca_coll_hierarch_verbose_param ) {
printf("%s:%d: executing hierarchical seg3 bcast with cnt=%d root=%d segsize=%d\n",
comm->c_name, rank, count, root, segsize );
}
/*
* This function returns the local leader communicator
* which *always* contains the root of this operation.
* This might involve creating a new communicator. This is
* also the reason, that *every* process in comm has to call
* this function
*/
llcomm = mca_coll_hierarch_get_llcomm ( root, hierarch_module, &llroot, &lroot);
ompi_datatype_type_size ( datatype, &typesize);
ompi_datatype_get_extent ( datatype, &ub, &typeext);
/* Determine number of segments and number of elements per segment */
if ((typesize > 0) && (segsize % typesize != 0)) {
/* segment size must be a multiple of typesize */
segsize = typesize * (segsize / typesize);
}
if ((segsize == 0) || (count == 0) || (typesize == 0)) {
segcount = count;
num_segments = 1;
} else {
segcount = segsize/typesize;
num_segments = count/segcount;
if ( (count % segcount) != 0 ) num_segments++;
if (num_segments == 1) segcount = count;
}
realsegsize = segcount*typeext;
remaining_count = segcount;
if ( MPI_COMM_NULL != lcomm ) {
lsize = ompi_comm_size ( lcomm );
lrank = ompi_comm_rank ( lcomm );
sreq1 = (ompi_request_t **)malloc ( lsize * sizeof(ompi_request_t *));
if ( NULL == sreq1 ) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
for(i=0; i<lsize; i++) {
sreq1[i] = MPI_REQUEST_NULL;
}
}
if ( MPI_COMM_NULL != llcomm ) {
llsize = ompi_comm_size (llcomm);
llrank = ompi_comm_rank ( llcomm );
sreq = hierarch_module->hier_reqs;
for(i=0; i<llsize; i++) {
sreq[i] = MPI_REQUEST_NULL;
}
}
/* Broadcasting the first segment in the upper level*/
if ( MPI_UNDEFINED != llroot ) {
ret = llcomm->c_coll.coll_bcast(tmpbuf, remaining_count, datatype,
llroot, llcomm,
llcomm->c_coll.coll_bcast_module);
if ( OMPI_SUCCESS != ret ) {
goto exit;
}
}
for (segindex = 1; segindex < num_segments; segindex++) {
/* determine how many elements are being sent in this round */
if( segindex == (num_segments - 1) ) {
remaining_count = count - segindex*segcount;
}
tmpbuf += realsegsize;
/* Broadcasting the next segment in the upper level*/
if ( MPI_COMM_NULL != llcomm ) {
if(llrank == llroot) {
for(i = 0; i < llsize; i++) {
if( i != llroot) {
ret = MCA_PML_CALL(isend(tmpbuf, remaining_count, datatype, i,
MCA_COLL_BASE_TAG_BCAST,
MCA_PML_BASE_SEND_STANDARD,
llcomm, (sreq+i) ));
if ( OMPI_SUCCESS != ret ) {
goto exit;
}
}
}
}
else {
ret = MCA_PML_CALL(irecv(tmpbuf, remaining_count, datatype, llroot,
MCA_COLL_BASE_TAG_BCAST,
llcomm, &rreq ));
if ( OMPI_SUCCESS != ret ) {
goto exit;
}
}
}
/* broadcasting the before segment among the lower level processes
* once the local leaders got the data from the root, they can distribute
* it to the processes in their local, low-level communicator.
*/
if ( MPI_COMM_NULL != lcomm ) {
if( lrank == lroot) {
for( i = 0; i < lsize; i++) {
if( i != lroot) {
ret = MCA_PML_CALL(isend(tmpbuf-realsegsize, segcount, datatype, i,
MCA_COLL_BASE_TAG_BCAST,
MCA_PML_BASE_SEND_STANDARD,
lcomm, (sreq1+i) ));
if ( OMPI_SUCCESS != ret ) {
goto exit;
}
}
}
}
else {
ret = MCA_PML_CALL(irecv(tmpbuf-realsegsize, segcount, datatype, lroot,
MCA_COLL_BASE_TAG_BCAST , lcomm, &rreq1 ));
if ( OMPI_SUCCESS != ret ) {
goto exit;
}
}
}
/* Wait for the upper level bcast to complete*/
if ( MPI_COMM_NULL != llcomm ) {
if ( llrank == llroot ) {
ret = ompi_request_wait_all(llsize, sreq, MPI_STATUSES_IGNORE);
if ( OMPI_SUCCESS != ret ) {
goto exit;
}
}
else {
ret = ompi_request_wait_all ( 1, &rreq, MPI_STATUS_IGNORE );
if ( OMPI_SUCCESS != ret ) {
goto exit;
}
}
}
/*Wait for the lower level bcast to complete */
if ( MPI_COMM_NULL != lcomm ) {
if ( lrank == lroot ) {
ret = ompi_request_wait_all(lsize, sreq1, MPI_STATUSES_IGNORE);
if ( OMPI_SUCCESS != ret ) {
goto exit;
}
}
else {
ret = ompi_request_wait_all( 1, &rreq1, MPI_STATUS_IGNORE);
if ( OMPI_SUCCESS != ret ) {
goto exit;
}
}
}
}
/*Bcasting the last segment among the lower level processes
* once the local leaders got the data from the root, they can distribute
* it to the processes in their local, low-level communicator.
*/
if ( MPI_COMM_NULL != lcomm ) {
ret = lcomm->c_coll.coll_bcast(tmpbuf, remaining_count, datatype,
lroot, lcomm,
lcomm->c_coll.coll_bcast_module);
}
exit:
if ( NULL != sreq1 ) {
free ( sreq1 );
}
return ret;
}
/*
* alltoallv_inter
*
* Function: - MPI_Alltoallv
* Accepts: - same as MPI_Alltoallv()
* Returns: - MPI_SUCCESS or an MPI error code
*/
int
mca_coll_basic_alltoallv_inter(const void *sbuf, const int *scounts, const int *sdisps,
struct ompi_datatype_t *sdtype, void *rbuf,
const int *rcounts, const int *rdisps,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int i;
int rsize;
int err;
char *psnd;
char *prcv;
size_t nreqs;
MPI_Aint sndextent;
MPI_Aint rcvextent;
mca_coll_basic_module_t *basic_module = (mca_coll_basic_module_t*) module;
ompi_request_t **preq = basic_module->mccb_reqs;
/* Initialize. */
rsize = ompi_comm_remote_size(comm);
ompi_datatype_type_extent(sdtype, &sndextent);
ompi_datatype_type_extent(rdtype, &rcvextent);
/* Initiate all send/recv to/from others. */
nreqs = rsize * 2;
/* Post all receives first */
/* A simple optimization: do not send and recv msgs of length zero */
for (i = 0; i < rsize; ++i) {
prcv = ((char *) rbuf) + (rdisps[i] * rcvextent);
if (rcounts[i] > 0) {
err = MCA_PML_CALL(irecv(prcv, rcounts[i], rdtype,
i, MCA_COLL_BASE_TAG_ALLTOALLV, comm,
&preq[i]));
if (MPI_SUCCESS != err) {
return err;
}
} else {
preq[i] = MPI_REQUEST_NULL;
}
}
/* Now post all sends */
for (i = 0; i < rsize; ++i) {
psnd = ((char *) sbuf) + (sdisps[i] * sndextent);
if (scounts[i] > 0) {
err = MCA_PML_CALL(isend(psnd, scounts[i], sdtype,
i, MCA_COLL_BASE_TAG_ALLTOALLV,
MCA_PML_BASE_SEND_STANDARD, comm,
&preq[rsize + i]));
if (MPI_SUCCESS != err) {
return err;
}
} else {
preq[rsize + i] = MPI_REQUEST_NULL;
}
}
err = ompi_request_wait_all(nreqs, preq, MPI_STATUSES_IGNORE);
/* All done */
return err;
}
/*
* Linear functions are copied from the basic coll module. For
* some small number of nodes and/or small data sizes they are just as
* fast as tuned/tree based segmenting operations and as such may be
* selected by the decision functions. These are copied into this module
* due to the way we select modules in V1. i.e. in V2 we will handle this
* differently and so will not have to duplicate code.
* GEF Oct05 after asking Jeff.
*/
int
ompi_coll_tuned_alltoallv_intra_basic_linear(void *sbuf, int *scounts, int *sdisps,
struct ompi_datatype_t *sdtype,
void *rbuf, int *rcounts, int *rdisps,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int i, size, rank, err, nreqs;
char *psnd, *prcv;
ptrdiff_t sext, rext;
MPI_Request *preq;
mca_coll_tuned_module_t *tuned_module = (mca_coll_tuned_module_t*) module;
mca_coll_tuned_comm_t *data = tuned_module->tuned_data;
if (MPI_IN_PLACE == sbuf) {
return mca_coll_tuned_alltoallv_intra_basic_inplace (rbuf, rcounts, rdisps,
rdtype, comm, module);
}
size = ompi_comm_size(comm);
rank = ompi_comm_rank(comm);
OPAL_OUTPUT((ompi_coll_tuned_stream,
"coll:tuned:alltoallv_intra_basic_linear rank %d", rank));
ompi_datatype_type_extent(sdtype, &sext);
ompi_datatype_type_extent(rdtype, &rext);
/* Simple optimization - handle send to self first */
psnd = ((char *) sbuf) + (ptrdiff_t)sdisps[rank] * sext;
prcv = ((char *) rbuf) + (ptrdiff_t)rdisps[rank] * rext;
if (0 != scounts[rank]) {
err = ompi_datatype_sndrcv(psnd, scounts[rank], sdtype,
prcv, rcounts[rank], rdtype);
if (MPI_SUCCESS != err) {
return err;
}
}
/* If only one process, we're done. */
if (1 == size) {
return MPI_SUCCESS;
}
/* Now, initiate all send/recv to/from others. */
nreqs = 0;
preq = data->mcct_reqs;
/* Post all receives first */
for (i = 0; i < size; ++i) {
if (i == rank || 0 == rcounts[i]) {
continue;
}
prcv = ((char *) rbuf) + (ptrdiff_t)rdisps[i] * rext;
err = MCA_PML_CALL(irecv_init(prcv, rcounts[i], rdtype,
i, MCA_COLL_BASE_TAG_ALLTOALLV, comm,
preq++));
++nreqs;
if (MPI_SUCCESS != err) {
ompi_coll_tuned_free_reqs(data->mcct_reqs, nreqs);
return err;
}
}
/* Now post all sends */
for (i = 0; i < size; ++i) {
if (i == rank || 0 == scounts[i]) {
continue;
}
psnd = ((char *) sbuf) + (ptrdiff_t)sdisps[i] * sext;
err = MCA_PML_CALL(isend_init(psnd, scounts[i], sdtype,
i, MCA_COLL_BASE_TAG_ALLTOALLV,
MCA_PML_BASE_SEND_STANDARD, comm,
preq++));
++nreqs;
if (MPI_SUCCESS != err) {
ompi_coll_tuned_free_reqs(data->mcct_reqs, nreqs);
return err;
}
}
/* Start your engines. This will never return an error. */
MCA_PML_CALL(start(nreqs, data->mcct_reqs));
/* Wait for them all. If there's an error, note that we don't care
* what the error was -- just that there *was* an error. The PML
* will finish all requests, even if one or more of them fail.
* i.e., by the end of this call, all the requests are free-able.
* So free them anyway -- even if there was an error, and return the
* error after we free everything. */
err = ompi_request_wait_all(nreqs, data->mcct_reqs,
MPI_STATUSES_IGNORE);
/* Free the requests. */
ompi_coll_tuned_free_reqs(data->mcct_reqs, nreqs);
return err;
}
static int
mca_coll_tuned_alltoallv_intra_basic_inplace(void *rbuf, const int *rcounts, const int *rdisps,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
mca_coll_tuned_module_t *tuned_module = (mca_coll_tuned_module_t*) module;
int i, j, size, rank, err=MPI_SUCCESS;
MPI_Request *preq;
char *tmp_buffer;
size_t max_size, rdtype_size;
ptrdiff_t ext;
/* Initialize. */
size = ompi_comm_size(comm);
rank = ompi_comm_rank(comm);
ompi_datatype_type_size(rdtype, &rdtype_size);
/* If only one process, we're done. */
if (1 == size || 0 == rdtype_size) {
return MPI_SUCCESS;
}
/* Find the largest receive amount */
ompi_datatype_type_extent (rdtype, &ext);
for (i = 0, max_size = 0 ; i < size ; ++i) {
size_t size = ext * rcounts[i];
max_size = size > max_size ? size : max_size;
}
/* Allocate a temporary buffer */
tmp_buffer = calloc (max_size, 1);
if (NULL == tmp_buffer) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* in-place alltoallv slow algorithm (but works) */
for (i = 0 ; i < size ; ++i) {
for (j = i+1 ; j < size ; ++j) {
/* Initiate all send/recv to/from others. */
preq = tuned_module->tuned_data->mcct_reqs;
if (i == rank && rcounts[j]) {
/* Copy the data into the temporary buffer */
err = ompi_datatype_copy_content_same_ddt (rdtype, rcounts[j],
tmp_buffer, (char *) rbuf + rdisps[j] * ext);
if (MPI_SUCCESS != err) { goto error_hndl; }
/* Exchange data with the peer */
err = MCA_PML_CALL(irecv ((char *) rbuf + rdisps[j] * ext, rcounts[j], rdtype,
j, MCA_COLL_BASE_TAG_ALLTOALLV, comm, preq++));
if (MPI_SUCCESS != err) { goto error_hndl; }
err = MCA_PML_CALL(isend ((void *) tmp_buffer, rcounts[j], rdtype,
j, MCA_COLL_BASE_TAG_ALLTOALLV, MCA_PML_BASE_SEND_STANDARD,
comm, preq++));
if (MPI_SUCCESS != err) { goto error_hndl; }
} else if (j == rank && rcounts[i]) {
/* Copy the data into the temporary buffer */
err = ompi_datatype_copy_content_same_ddt (rdtype, rcounts[i],
tmp_buffer, (char *) rbuf + rdisps[i] * ext);
if (MPI_SUCCESS != err) { goto error_hndl; }
/* Exchange data with the peer */
err = MCA_PML_CALL(irecv ((char *) rbuf + rdisps[i] * ext, rcounts[i], rdtype,
i, MCA_COLL_BASE_TAG_ALLTOALLV, comm, preq++));
if (MPI_SUCCESS != err) { goto error_hndl; }
err = MCA_PML_CALL(isend ((void *) tmp_buffer, rcounts[i], rdtype,
i, MCA_COLL_BASE_TAG_ALLTOALLV, MCA_PML_BASE_SEND_STANDARD,
comm, preq++));
if (MPI_SUCCESS != err) { goto error_hndl; }
} else {
continue;
}
/* Wait for the requests to complete */
err = ompi_request_wait_all (2, tuned_module->tuned_data->mcct_reqs, MPI_STATUSES_IGNORE);
if (MPI_SUCCESS != err) { goto error_hndl; }
/* Free the requests. */
mca_coll_tuned_free_reqs(tuned_module->tuned_data->mcct_reqs, 2);
}
}
error_hndl:
/* Free the temporary buffer */
free (tmp_buffer);
/* All done */
return err;
}
/**
* Bcast - subgroup in communicator
* This is a very simple algorithm - binary tree, transmitting the full
* message at each step.
*/
OMPI_DECLSPEC int comm_bcast_pml(void *buffer, int root, int count,
ompi_datatype_t *dtype, int my_rank_in_group,
int n_peers, int *ranks_in_comm,ompi_communicator_t *comm)
{
/* local variables */
int rc=OMPI_SUCCESS,msg_cnt,i;
ompi_request_t *requests[2];
int node_rank, peer_rank;
netpatterns_tree_node_t node_data;
/*
* shift rank to root==0 tree
*/
node_rank=(my_rank_in_group-root+n_peers)%n_peers;
/*
* compute my communication pattern - binary tree
*/
rc=netpatterns_setup_narray_tree(2, node_rank, n_peers,
&node_data);
if( OMPI_SUCCESS != rc ) {
goto Error;
}
/* 1 process special case */
if(1 == n_peers) {
return OMPI_SUCCESS;
}
/* if I have parents - wait on the data to arrive */
if(node_data.n_parents) {
/* I will have only 1 parent */
peer_rank=node_data.parent_rank;
peer_rank=(peer_rank+root)%n_peers;
/* translate back to actual rank */
rc=MCA_PML_CALL(recv(buffer, count,dtype,peer_rank,
-OMPI_COMMON_TAG_BCAST, comm, MPI_STATUSES_IGNORE));
if( 0 > rc ) {
goto Error;
}
}
/* send the data to my children */
msg_cnt=0;
for(i=0 ; i < node_data.n_children ; i++ ) {
peer_rank=node_data.children_ranks[i];
peer_rank=(peer_rank+root)%n_peers;
rc=MCA_PML_CALL(isend(buffer,
count,dtype,peer_rank,
-OMPI_COMMON_TAG_BCAST,MCA_PML_BASE_SEND_STANDARD,
comm,&(requests[msg_cnt])));
if( 0 > rc ) {
goto Error;
}
msg_cnt++;
}
/* wait for send completion */
if(msg_cnt) {
/* wait on send and receive completion */
ompi_request_wait_all(msg_cnt,requests,MPI_STATUSES_IGNORE);
}
/* return */
return OMPI_SUCCESS;
Error:
return rc;
}
static int two_phase_exchage_data(mca_io_ompio_file_t *fh,
void *buf,
char *write_buf,
struct iovec *offset_length,
int *send_size,int *start_pos,
int *recv_size,
OMPI_MPI_OFFSET_TYPE off,
OMPI_MPI_OFFSET_TYPE size, int *count,
int *partial_recv, int *sent_to_proc,
int contig_access_count,
OMPI_MPI_OFFSET_TYPE min_st_offset,
OMPI_MPI_OFFSET_TYPE fd_size,
OMPI_MPI_OFFSET_TYPE *fd_start,
OMPI_MPI_OFFSET_TYPE *fd_end,
Flatlist_node *flat_buf,
mca_io_ompio_access_array_t *others_req,
int *send_buf_idx, int *curr_to_proc,
int *done_to_proc, int iter,
int *buf_idx,MPI_Aint buftype_extent,
int striping_unit, int *aggregator_list,
int *hole){
int *tmp_len=NULL, sum, *srt_len=NULL, nprocs_recv, nprocs_send, k,i,j;
int ret=OMPI_SUCCESS;
MPI_Request *requests=NULL, *send_req=NULL;
MPI_Datatype *recv_types=NULL;
OMPI_MPI_OFFSET_TYPE *srt_off=NULL;
char **send_buf = NULL;
ret = fh->f_comm->c_coll.coll_alltoall (recv_size,
1,
MPI_INT,
send_size,
1,
MPI_INT,
fh->f_comm,
fh->f_comm->c_coll.coll_alltoall_module);
if ( OMPI_SUCCESS != ret ){
return ret;
}
nprocs_recv = 0;
for (i=0;i<fh->f_size;i++){
if (recv_size[i]){
nprocs_recv++;
}
}
recv_types = (MPI_Datatype *)
malloc (( nprocs_recv + 1 ) * sizeof(MPI_Datatype *));
if ( NULL == recv_types ){
return OMPI_ERR_OUT_OF_RESOURCE;
}
tmp_len = (int *) malloc(fh->f_size*sizeof(int));
if ( NULL == tmp_len ) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
j = 0;
for (i=0;i<fh->f_size;i++){
if (recv_size[i]) {
if (partial_recv[i]) {
k = start_pos[i] + count[i] - 1;
tmp_len[i] = others_req[i].lens[k];
others_req[i].lens[k] = partial_recv[i];
}
MPI_Type_hindexed(count[i],
&(others_req[i].lens[start_pos[i]]),
&(others_req[i].mem_ptrs[start_pos[i]]),
MPI_BYTE, recv_types+j);
MPI_Type_commit(recv_types+j);
j++;
}
}
sum = 0;
for (i=0;i<fh->f_size;i++) sum += count[i];
srt_off = (OMPI_MPI_OFFSET_TYPE *)
malloc((sum+1)*sizeof(OMPI_MPI_OFFSET_TYPE));
if ( NULL == srt_off ){
return OMPI_ERR_OUT_OF_RESOURCE;
}
srt_len = (int *) malloc((sum+1)*sizeof(int));
if ( NULL == srt_len ) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
two_phase_heap_merge(others_req, count, srt_off, srt_len, start_pos, fh->f_size,fh->f_rank, nprocs_recv, sum);
for (i=0; i<fh->f_size; i++)
if (partial_recv[i]) {
k = start_pos[i] + count[i] - 1;
others_req[i].lens[k] = tmp_len[i];
}
if ( NULL != tmp_len ){
free(tmp_len);
}
*hole = 0;
if (off != srt_off[0]){
*hole = 1;
}
else{
for (i=1;i<sum;i++){
if (srt_off[i] <= srt_off[0] + srt_len[0]){
int new_len = srt_off[i] + srt_len[i] - srt_off[0];
if(new_len > srt_len[0])
srt_len[0] = new_len;
}
else
break;
}
if (i < sum || size != srt_len[0])
*hole = 1;
}
if ( NULL != srt_off ){
free(srt_off);
}
if ( NULL != srt_len ){
free(srt_len);
}
if (nprocs_recv){
if (*hole){
if (off > 0){
fh->f_io_array = (mca_io_ompio_io_array_t *)malloc
(sizeof(mca_io_ompio_io_array_t));
if (NULL == fh->f_io_array) {
opal_output(1, "OUT OF MEMORY\n");
return OMPI_ERR_OUT_OF_RESOURCE;
}
fh->f_io_array[0].offset =(IOVBASE_TYPE *)(intptr_t)off;
fh->f_num_of_io_entries = 1;
fh->f_io_array[0].length = size;
fh->f_io_array[0].memory_address = write_buf;
if (fh->f_num_of_io_entries){
if (OMPI_SUCCESS != fh->f_fbtl->fbtl_preadv (fh, NULL)) {
opal_output(1, "READ FAILED\n");
return OMPI_ERROR;
}
}
}
fh->f_num_of_io_entries = 0;
if (NULL != fh->f_io_array) {
free (fh->f_io_array);
fh->f_io_array = NULL;
}
}
}
nprocs_send = 0;
for (i=0; i <fh->f_size; i++) if (send_size[i]) nprocs_send++;
#if DEBUG_ON
printf("%d : nprocs_send : %d\n", fh->f_rank,nprocs_send);
#endif
requests = (MPI_Request *)
malloc((nprocs_send+nprocs_recv+1)*sizeof(MPI_Request));
if ( NULL == requests ){
return OMPI_ERR_OUT_OF_RESOURCE;
}
j = 0;
for (i=0; i<fh->f_size; i++) {
if (recv_size[i]) {
ret = MCA_PML_CALL(irecv(MPI_BOTTOM,
1,
recv_types[j],
i,
fh->f_rank+i+100*iter,
fh->f_comm,
requests+j));
if ( OMPI_SUCCESS != ret ){
return ret;
}
j++;
}
}
send_req = requests + nprocs_recv;
if (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY) {
j = 0;
for (i=0; i <fh->f_size; i++)
if (send_size[i]) {
ret = MCA_PML_CALL(isend(((char *) buf) + buf_idx[i],
send_size[i],
MPI_BYTE,
i,
fh->f_rank+i+100*iter,
MCA_PML_BASE_SEND_STANDARD,
fh->f_comm,
send_req+j));
if ( OMPI_SUCCESS != ret ){
return ret;
}
j++;
buf_idx[i] += send_size[i];
}
}
else if(nprocs_send && (!(fh->f_flags & OMPIO_CONTIGUOUS_MEMORY))){
send_buf = (char **) malloc(fh->f_size*sizeof(char*));
if ( NULL == send_buf ){
return OMPI_ERR_OUT_OF_RESOURCE;
}
for (i=0; i < fh->f_size; i++){
if (send_size[i]) {
send_buf[i] = (char *) malloc(send_size[i]);
if ( NULL == send_buf[i] ){
return OMPI_ERR_OUT_OF_RESOURCE;
}
}
}
ret = two_phase_fill_send_buffer(fh, buf,flat_buf, send_buf,
offset_length, send_size,
send_req,sent_to_proc,
contig_access_count,
min_st_offset, fd_size,
fd_start, fd_end, send_buf_idx,
curr_to_proc, done_to_proc,
iter, buftype_extent, striping_unit,
aggregator_list);
if ( OMPI_SUCCESS != ret ){
return ret;
}
}
for (i=0; i<nprocs_recv; i++) MPI_Type_free(recv_types+i);
free(recv_types);
ret = ompi_request_wait_all (nprocs_send+nprocs_recv,
requests,
MPI_STATUS_IGNORE);
if ( NULL != requests ){
free(requests);
}
return ret;
}
/*
* gather_intra_linear_sync
*
* Function: - synchronized gather operation with
* Accepts: - same arguments as MPI_Gather(), first segment size
* Returns: - MPI_SUCCESS or error code
*/
int
ompi_coll_base_gather_intra_linear_sync(const void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
void *rbuf, int rcount,
struct ompi_datatype_t *rdtype,
int root,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module,
int first_segment_size)
{
int i, ret, line, rank, size, first_segment_count;
ompi_request_t **reqs = NULL;
MPI_Aint extent, lb;
size_t typelng;
size = ompi_comm_size(comm);
rank = ompi_comm_rank(comm);
OPAL_OUTPUT((ompi_coll_base_framework.framework_output,
"ompi_coll_base_gather_intra_linear_sync rank %d, segment %d", rank, first_segment_size));
if (rank != root) {
/* Non-root processes:
- receive zero byte message from the root,
- send the first segment of the data synchronously,
- send the second segment of the data.
*/
ompi_datatype_type_size(sdtype, &typelng);
ompi_datatype_get_extent(sdtype, &lb, &extent);
first_segment_count = scount;
COLL_BASE_COMPUTED_SEGCOUNT( (size_t) first_segment_size, typelng,
first_segment_count );
ret = MCA_PML_CALL(recv(rbuf, 0, MPI_BYTE, root,
MCA_COLL_BASE_TAG_GATHER,
comm, MPI_STATUS_IGNORE));
if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
ret = MCA_PML_CALL(send(sbuf, first_segment_count, sdtype, root,
MCA_COLL_BASE_TAG_GATHER,
MCA_PML_BASE_SEND_STANDARD, comm));
if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
ret = MCA_PML_CALL(send((char*)sbuf + extent * first_segment_count,
(scount - first_segment_count), sdtype,
root, MCA_COLL_BASE_TAG_GATHER,
MCA_PML_BASE_SEND_STANDARD, comm));
if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
} else {
/* Root process,
- For every non-root node:
- post irecv for the first segment of the message
- send zero byte message to signal node to send the message
- post irecv for the second segment of the message
- wait for the first segment to complete
- Copy local data if necessary
- Waitall for all the second segments to complete.
*/
char *ptmp;
ompi_request_t *first_segment_req;
reqs = coll_base_comm_get_reqs(module->base_data, size);
if (NULL == reqs) { ret = -1; line = __LINE__; goto error_hndl; }
ompi_datatype_type_size(rdtype, &typelng);
ompi_datatype_get_extent(rdtype, &lb, &extent);
first_segment_count = rcount;
COLL_BASE_COMPUTED_SEGCOUNT( (size_t)first_segment_size, typelng,
first_segment_count );
ptmp = (char *) rbuf;
for (i = 0; i < size; ++i) {
if (i == rank) {
/* skip myself */
reqs[i] = MPI_REQUEST_NULL;
continue;
}
/* irecv for the first segment from i */
ptmp = (char*)rbuf + (ptrdiff_t)i * (ptrdiff_t)rcount * extent;
ret = MCA_PML_CALL(irecv(ptmp, first_segment_count, rdtype, i,
MCA_COLL_BASE_TAG_GATHER, comm,
&first_segment_req));
if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
/* send sync message */
ret = MCA_PML_CALL(send(rbuf, 0, MPI_BYTE, i,
MCA_COLL_BASE_TAG_GATHER,
MCA_PML_BASE_SEND_STANDARD, comm));
if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
/* irecv for the second segment */
ptmp = (char*)rbuf + ((ptrdiff_t)i * (ptrdiff_t)rcount + first_segment_count) * extent;
ret = MCA_PML_CALL(irecv(ptmp, (rcount - first_segment_count),
rdtype, i, MCA_COLL_BASE_TAG_GATHER, comm,
&reqs[i]));
if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
/* wait on the first segment to complete */
ret = ompi_request_wait(&first_segment_req, MPI_STATUS_IGNORE);
if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
}
/* copy local data if necessary */
if (MPI_IN_PLACE != sbuf) {
ret = ompi_datatype_sndrcv((void *)sbuf, scount, sdtype,
(char*)rbuf + (ptrdiff_t)rank * (ptrdiff_t)rcount * extent,
rcount, rdtype);
if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
}
/* wait all second segments to complete */
ret = ompi_request_wait_all(size, reqs, MPI_STATUSES_IGNORE);
if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
}
/* All done */
return MPI_SUCCESS;
error_hndl:
if (NULL != reqs) {
ompi_coll_base_free_reqs(reqs, size);
}
OPAL_OUTPUT (( ompi_coll_base_framework.framework_output,
"ERROR_HNDL: node %d file %s line %d error %d\n",
rank, __FILE__, line, ret ));
(void)line; // silence compiler warning
return ret;
}
static int
mca_coll_basic_neighbor_alltoallv_cart(const void *sbuf, const int scounts[], const int sdisps[],
struct ompi_datatype_t *sdtype, void *rbuf, const int rcounts[],
const int rdisps[], struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm, mca_coll_base_module_t *module)
{
mca_coll_basic_module_t *basic_module = (mca_coll_basic_module_t *) module;
const mca_topo_base_comm_cart_2_1_0_t *cart = comm->c_topo->mtc.cart;
const int rank = ompi_comm_rank (comm);
int rc = MPI_SUCCESS, dim, i, nreqs;
ptrdiff_t lb, rdextent, sdextent;
ompi_request_t **reqs;
ompi_datatype_get_extent(rdtype, &lb, &rdextent);
ompi_datatype_get_extent(sdtype, &lb, &sdextent);
/* post receives first */
for (dim = 0, nreqs = 0, i = 0, reqs = basic_module->mccb_reqs ; dim < cart->ndims ; ++dim, i += 2) {
int srank = MPI_PROC_NULL, drank = MPI_PROC_NULL;
if (cart->dims[dim] > 1) {
mca_topo_base_cart_shift (comm, dim, 1, &srank, &drank);
} else if (1 == cart->dims[dim] && cart->periods[dim]) {
srank = drank = rank;
}
if (MPI_PROC_NULL != srank) {
rc = MCA_PML_CALL(irecv((char *) rbuf + rdisps[i] * rdextent, rcounts[i], rdtype, srank,
MCA_COLL_BASE_TAG_ALLTOALL, comm, reqs++));
if (OMPI_SUCCESS != rc) break;
nreqs++;
}
if (MPI_PROC_NULL != drank) {
rc = MCA_PML_CALL(irecv((char *) rbuf + rdisps[i+1] * rdextent, rcounts[i+1], rdtype, drank,
MCA_COLL_BASE_TAG_ALLTOALL, comm, reqs++));
if (OMPI_SUCCESS != rc) break;
nreqs++;
}
}
if (OMPI_SUCCESS != rc) {
/* should probably try to clean up here */
return rc;
}
for (dim = 0, i = 0 ; dim < cart->ndims ; ++dim, i += 2) {
int srank = MPI_PROC_NULL, drank = MPI_PROC_NULL;
if (cart->dims[dim] > 1) {
mca_topo_base_cart_shift (comm, dim, 1, &srank, &drank);
} else if (1 == cart->dims[dim] && cart->periods[dim]) {
srank = drank = rank;
}
if (MPI_PROC_NULL != srank) {
/* remove cast from const when the pml layer is updated to take a const for the send buffer */
rc = MCA_PML_CALL(isend((char *) sbuf + sdisps[i] * sdextent, scounts[i], sdtype, srank,
MCA_COLL_BASE_TAG_ALLTOALL, MCA_PML_BASE_SEND_STANDARD, comm, reqs++));
if (OMPI_SUCCESS != rc) break;
nreqs++;
}
if (MPI_PROC_NULL != drank) {
rc = MCA_PML_CALL(isend((char *) sbuf + sdisps[i+1] * sdextent, scounts[i+1], sdtype, drank,
MCA_COLL_BASE_TAG_ALLTOALL, MCA_PML_BASE_SEND_STANDARD, comm, reqs++));
if (OMPI_SUCCESS != rc) break;
nreqs++;
}
}
if (OMPI_SUCCESS != rc) {
/* should probably try to clean up here */
return rc;
}
return ompi_request_wait_all (nreqs, basic_module->mccb_reqs, MPI_STATUSES_IGNORE);
}
/**
* This is a generic implementation of the reduce protocol. It used the tree
* provided as an argument and execute all operations using a segment of
* count times a datatype.
* For the last communication it will update the count in order to limit
* the number of datatype to the original count (original_count)
*
* Note that for non-commutative operations we cannot save memory copy
* for the first block: thus we must copy sendbuf to accumbuf on intermediate
* to keep the optimized loop happy.
*/
int ompi_coll_base_reduce_generic( const void* sendbuf, void* recvbuf, int original_count,
ompi_datatype_t* datatype, ompi_op_t* op,
int root, ompi_communicator_t* comm,
mca_coll_base_module_t *module,
ompi_coll_tree_t* tree, int count_by_segment,
int max_outstanding_reqs )
{
char *inbuf[2] = {NULL, NULL}, *inbuf_free[2] = {NULL, NULL};
char *accumbuf = NULL, *accumbuf_free = NULL;
char *local_op_buffer = NULL, *sendtmpbuf = NULL;
ptrdiff_t extent, size, gap, segment_increment;
ompi_request_t **sreq = NULL, *reqs[2] = {MPI_REQUEST_NULL, MPI_REQUEST_NULL};
int num_segments, line, ret, segindex, i, rank;
int recvcount, prevcount, inbi;
/**
* Determine number of segments and number of elements
* sent per operation
*/
ompi_datatype_type_extent( datatype, &extent );
num_segments = (int)(((size_t)original_count + (size_t)count_by_segment - (size_t)1) / (size_t)count_by_segment);
segment_increment = (ptrdiff_t)count_by_segment * extent;
sendtmpbuf = (char*) sendbuf;
if( sendbuf == MPI_IN_PLACE ) {
sendtmpbuf = (char *)recvbuf;
}
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, "coll:base:reduce_generic count %d, msg size %ld, segsize %ld, max_requests %d",
original_count, (unsigned long)((ptrdiff_t)num_segments * (ptrdiff_t)segment_increment),
(unsigned long)segment_increment, max_outstanding_reqs));
rank = ompi_comm_rank(comm);
/* non-leaf nodes - wait for children to send me data & forward up
(if needed) */
if( tree->tree_nextsize > 0 ) {
ptrdiff_t real_segment_size;
/* handle non existant recv buffer (i.e. its NULL) and
protect the recv buffer on non-root nodes */
accumbuf = (char*)recvbuf;
if( (NULL == accumbuf) || (root != rank) ) {
/* Allocate temporary accumulator buffer. */
size = opal_datatype_span(&datatype->super, original_count, &gap);
accumbuf_free = (char*)malloc(size);
if (accumbuf_free == NULL) {
line = __LINE__; ret = -1; goto error_hndl;
}
accumbuf = accumbuf_free - gap;
}
/* If this is a non-commutative operation we must copy
sendbuf to the accumbuf, in order to simplfy the loops */
if (!ompi_op_is_commute(op)) {
ompi_datatype_copy_content_same_ddt(datatype, original_count,
(char*)accumbuf,
(char*)sendtmpbuf);
}
/* Allocate two buffers for incoming segments */
real_segment_size = opal_datatype_span(&datatype->super, count_by_segment, &gap);
inbuf_free[0] = (char*) malloc(real_segment_size);
if( inbuf_free[0] == NULL ) {
line = __LINE__; ret = -1; goto error_hndl;
}
inbuf[0] = inbuf_free[0] - gap;
/* if there is chance to overlap communication -
allocate second buffer */
if( (num_segments > 1) || (tree->tree_nextsize > 1) ) {
inbuf_free[1] = (char*) malloc(real_segment_size);
if( inbuf_free[1] == NULL ) {
line = __LINE__; ret = -1; goto error_hndl;
}
inbuf[1] = inbuf_free[1] - gap;
}
/* reset input buffer index and receive count */
inbi = 0;
recvcount = 0;
/* for each segment */
for( segindex = 0; segindex <= num_segments; segindex++ ) {
prevcount = recvcount;
/* recvcount - number of elements in current segment */
recvcount = count_by_segment;
if( segindex == (num_segments-1) )
recvcount = original_count - (ptrdiff_t)count_by_segment * (ptrdiff_t)segindex;
/* for each child */
for( i = 0; i < tree->tree_nextsize; i++ ) {
/**
* We try to overlap communication:
* either with next segment or with the next child
*/
/* post irecv for current segindex on current child */
if( segindex < num_segments ) {
void* local_recvbuf = inbuf[inbi];
if( 0 == i ) {
/* for the first step (1st child per segment) and
* commutative operations we might be able to irecv
* directly into the accumulate buffer so that we can
* reduce(op) this with our sendbuf in one step as
* ompi_op_reduce only has two buffer pointers,
* this avoids an extra memory copy.
*
* BUT if the operation is non-commutative or
* we are root and are USING MPI_IN_PLACE this is wrong!
*/
if( (ompi_op_is_commute(op)) &&
!((MPI_IN_PLACE == sendbuf) && (rank == tree->tree_root)) ) {
local_recvbuf = accumbuf + (ptrdiff_t)segindex * (ptrdiff_t)segment_increment;
}
}
ret = MCA_PML_CALL(irecv(local_recvbuf, recvcount, datatype,
tree->tree_next[i],
MCA_COLL_BASE_TAG_REDUCE, comm,
&reqs[inbi]));
if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl;}
}
/* wait for previous req to complete, if any.
if there are no requests reqs[inbi ^1] will be
MPI_REQUEST_NULL. */
/* wait on data from last child for previous segment */
ret = ompi_request_wait_all( 1, &reqs[inbi ^ 1],
MPI_STATUSES_IGNORE );
if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
local_op_buffer = inbuf[inbi ^ 1];
if( i > 0 ) {
/* our first operation is to combine our own [sendbuf] data
* with the data we recvd from down stream (but only
* the operation is commutative and if we are not root and
* not using MPI_IN_PLACE)
*/
if( 1 == i ) {
if( (ompi_op_is_commute(op)) &&
!((MPI_IN_PLACE == sendbuf) && (rank == tree->tree_root)) ) {
local_op_buffer = sendtmpbuf + (ptrdiff_t)segindex * (ptrdiff_t)segment_increment;
}
}
/* apply operation */
ompi_op_reduce(op, local_op_buffer,
accumbuf + (ptrdiff_t)segindex * (ptrdiff_t)segment_increment,
recvcount, datatype );
} else if ( segindex > 0 ) {
void* accumulator = accumbuf + (ptrdiff_t)(segindex-1) * (ptrdiff_t)segment_increment;
if( tree->tree_nextsize <= 1 ) {
if( (ompi_op_is_commute(op)) &&
!((MPI_IN_PLACE == sendbuf) && (rank == tree->tree_root)) ) {
local_op_buffer = sendtmpbuf + (ptrdiff_t)(segindex-1) * (ptrdiff_t)segment_increment;
}
}
ompi_op_reduce(op, local_op_buffer, accumulator, prevcount,
datatype );
/* all reduced on available data this step (i) complete,
* pass to the next process unless you are the root.
*/
if (rank != tree->tree_root) {
/* send combined/accumulated data to parent */
ret = MCA_PML_CALL( send( accumulator, prevcount,
datatype, tree->tree_prev,
MCA_COLL_BASE_TAG_REDUCE,
MCA_PML_BASE_SEND_STANDARD,
comm) );
if (ret != MPI_SUCCESS) {
line = __LINE__; goto error_hndl;
}
}
/* we stop when segindex = number of segments
(i.e. we do num_segment+1 steps for pipelining */
if (segindex == num_segments) break;
}
/* update input buffer index */
inbi = inbi ^ 1;
} /* end of for each child */
} /* end of for each segment */
/* clean up */
if( inbuf_free[0] != NULL) free(inbuf_free[0]);
if( inbuf_free[1] != NULL) free(inbuf_free[1]);
if( accumbuf_free != NULL ) free(accumbuf_free);
}
/* leaf nodes
Depending on the value of max_outstanding_reqs and
the number of segments we have two options:
- send all segments using blocking send to the parent, or
- avoid overflooding the parent nodes by limiting the number of
outstanding requests to max_oustanding_reqs.
TODO/POSSIBLE IMPROVEMENT: If there is a way to determine the eager size
for the current communication, synchronization should be used only
when the message/segment size is smaller than the eager size.
*/
else {
/* If the number of segments is less than a maximum number of oustanding
requests or there is no limit on the maximum number of outstanding
requests, we send data to the parent using blocking send */
if ((0 == max_outstanding_reqs) ||
(num_segments <= max_outstanding_reqs)) {
segindex = 0;
while ( original_count > 0) {
if (original_count < count_by_segment) {
count_by_segment = original_count;
}
ret = MCA_PML_CALL( send((char*)sendbuf +
(ptrdiff_t)segindex * (ptrdiff_t)segment_increment,
count_by_segment, datatype,
tree->tree_prev,
MCA_COLL_BASE_TAG_REDUCE,
MCA_PML_BASE_SEND_STANDARD,
comm) );
if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
segindex++;
original_count -= count_by_segment;
}
}
/* Otherwise, introduce flow control:
- post max_outstanding_reqs non-blocking synchronous send,
- for remaining segments
- wait for a ssend to complete, and post the next one.
- wait for all outstanding sends to complete.
*/
else {
int creq = 0;
sreq = coll_base_comm_get_reqs(module->base_data, max_outstanding_reqs);
if (NULL == sreq) { line = __LINE__; ret = -1; goto error_hndl; }
/* post first group of requests */
for (segindex = 0; segindex < max_outstanding_reqs; segindex++) {
ret = MCA_PML_CALL( isend((char*)sendbuf +
(ptrdiff_t)segindex * (ptrdiff_t)segment_increment,
count_by_segment, datatype,
tree->tree_prev,
MCA_COLL_BASE_TAG_REDUCE,
MCA_PML_BASE_SEND_SYNCHRONOUS, comm,
&sreq[segindex]) );
if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
original_count -= count_by_segment;
}
creq = 0;
while ( original_count > 0 ) {
/* wait on a posted request to complete */
ret = ompi_request_wait(&sreq[creq], MPI_STATUS_IGNORE);
if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
if( original_count < count_by_segment ) {
count_by_segment = original_count;
}
ret = MCA_PML_CALL( isend((char*)sendbuf +
(ptrdiff_t)segindex * (ptrdiff_t)segment_increment,
count_by_segment, datatype,
tree->tree_prev,
MCA_COLL_BASE_TAG_REDUCE,
MCA_PML_BASE_SEND_SYNCHRONOUS, comm,
&sreq[creq]) );
if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
creq = (creq + 1) % max_outstanding_reqs;
segindex++;
original_count -= count_by_segment;
}
/* Wait on the remaining request to complete */
ret = ompi_request_wait_all( max_outstanding_reqs, sreq,
MPI_STATUSES_IGNORE );
if (ret != MPI_SUCCESS) { line = __LINE__; goto error_hndl; }
}
}
return OMPI_SUCCESS;
error_hndl: /* error handler */
OPAL_OUTPUT (( ompi_coll_base_framework.framework_output,
"ERROR_HNDL: node %d file %s line %d error %d\n",
rank, __FILE__, line, ret ));
(void)line; // silence compiler warning
if( inbuf_free[0] != NULL ) free(inbuf_free[0]);
if( inbuf_free[1] != NULL ) free(inbuf_free[1]);
if( accumbuf_free != NULL ) free(accumbuf);
if( NULL != sreq ) {
ompi_coll_base_free_reqs(sreq, max_outstanding_reqs);
}
return ret;
}
/* MPI_IN_PLACE all to all algorithm. TODO: implement a better one. */
int
mca_coll_base_alltoall_intra_basic_inplace(const void *rbuf, int rcount,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
mca_coll_base_module_t *base_module = (mca_coll_base_module_t*) module;
int i, j, size, rank, err = MPI_SUCCESS, line;
MPI_Request *preq;
char *tmp_buffer;
size_t max_size;
ptrdiff_t ext;
/* Initialize. */
size = ompi_comm_size(comm);
rank = ompi_comm_rank(comm);
/* If only one process, we're done. */
if (1 == size) {
return MPI_SUCCESS;
}
/* Find the largest receive amount */
ompi_datatype_type_extent (rdtype, &ext);
max_size = ext * rcount;
/* Allocate a temporary buffer */
tmp_buffer = calloc (max_size, 1);
if (NULL == tmp_buffer) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* in-place alltoall slow algorithm (but works) */
for (i = 0 ; i < size ; ++i) {
for (j = i+1 ; j < size ; ++j) {
/* Initiate all send/recv to/from others. */
preq = coll_base_comm_get_reqs(base_module->base_data, size * 2);
if (i == rank) {
/* Copy the data into the temporary buffer */
err = ompi_datatype_copy_content_same_ddt (rdtype, rcount, tmp_buffer,
(char *) rbuf + j * max_size);
if (MPI_SUCCESS != err) { line = __LINE__; goto error_hndl; }
/* Exchange data with the peer */
err = MCA_PML_CALL(irecv ((char *) rbuf + max_size * j, rcount, rdtype,
j, MCA_COLL_BASE_TAG_ALLTOALL, comm, preq++));
if (MPI_SUCCESS != err) { line = __LINE__; goto error_hndl; }
err = MCA_PML_CALL(isend ((char *) tmp_buffer, rcount, rdtype,
j, MCA_COLL_BASE_TAG_ALLTOALL, MCA_PML_BASE_SEND_STANDARD,
comm, preq++));
if (MPI_SUCCESS != err) { line = __LINE__; goto error_hndl; }
} else if (j == rank) {
/* Copy the data into the temporary buffer */
err = ompi_datatype_copy_content_same_ddt (rdtype, rcount, tmp_buffer,
(char *) rbuf + i * max_size);
if (MPI_SUCCESS != err) { line = __LINE__; goto error_hndl; }
/* Exchange data with the peer */
err = MCA_PML_CALL(irecv ((char *) rbuf + max_size * i, rcount, rdtype,
i, MCA_COLL_BASE_TAG_ALLTOALL, comm, preq++));
if (MPI_SUCCESS != err) { line = __LINE__; goto error_hndl; }
err = MCA_PML_CALL(isend ((char *) tmp_buffer, rcount, rdtype,
i, MCA_COLL_BASE_TAG_ALLTOALL, MCA_PML_BASE_SEND_STANDARD,
comm, preq++));
if (MPI_SUCCESS != err) { line = __LINE__; goto error_hndl; }
} else {
continue;
}
/* Wait for the requests to complete */
err = ompi_request_wait_all (2, base_module->base_data->mcct_reqs, MPI_STATUSES_IGNORE);
if (MPI_SUCCESS != err) { line = __LINE__; goto error_hndl; }
}
}
error_hndl:
/* Free the temporary buffer */
free (tmp_buffer);
if( MPI_SUCCESS != err ) {
OPAL_OUTPUT((ompi_coll_base_framework.framework_output,
"%s:%4d\tError occurred %d, rank %2d", __FILE__, line, err,
rank));
ompi_coll_base_free_reqs(base_module->base_data->mcct_reqs, 2);
}
/* All done */
return err;
}
int
mca_fcoll_dynamic_file_read_all (mca_io_ompio_file_t *fh,
void *buf,
int count,
struct ompi_datatype_t *datatype,
ompi_status_public_t *status)
{
MPI_Aint position = 0;
MPI_Aint total_bytes = 0; /* total bytes to be read */
MPI_Aint bytes_to_read_in_cycle = 0; /* left to be read in a cycle*/
MPI_Aint bytes_per_cycle = 0; /* total read in each cycle by each process*/
int index = 0, ret=OMPI_SUCCESS;
int cycles = 0;
int i=0, j=0, l=0;
int n=0; /* current position in total_bytes_per_process array */
MPI_Aint bytes_remaining = 0; /* how many bytes have been read from the current
value from total_bytes_per_process */
int *sorted_file_offsets=NULL, entries_per_aggregator=0;
int bytes_received = 0;
int blocks = 0;
/* iovec structure and count of the buffer passed in */
uint32_t iov_count = 0;
struct iovec *decoded_iov = NULL;
int iov_index = 0;
size_t current_position = 0;
struct iovec *local_iov_array=NULL, *global_iov_array=NULL;
char *receive_buf = NULL;
MPI_Aint *memory_displacements=NULL;
/* global iovec at the readers that contain the iovecs created from
file_set_view */
uint32_t total_fview_count = 0;
int local_count = 0;
int *fview_count = NULL, *disp_index=NULL, *temp_disp_index=NULL;
int current_index=0, temp_index=0;
int **blocklen_per_process=NULL;
MPI_Aint **displs_per_process=NULL;
char *global_buf = NULL;
MPI_Aint global_count = 0;
local_io_array *file_offsets_for_agg=NULL;
/* array that contains the sorted indices of the global_iov */
int *sorted = NULL;
int *displs = NULL;
int dynamic_num_io_procs;
size_t max_data = 0;
int *bytes_per_process = NULL;
MPI_Aint *total_bytes_per_process = NULL;
ompi_datatype_t **sendtype = NULL;
MPI_Request *send_req=NULL, *recv_req=NULL;
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
double read_time = 0.0, start_read_time = 0.0, end_read_time = 0.0;
double rcomm_time = 0.0, start_rcomm_time = 0.0, end_rcomm_time = 0.0;
double read_exch = 0.0, start_rexch = 0.0, end_rexch = 0.0;
mca_io_ompio_print_entry nentry;
#endif
// if (opal_datatype_is_contiguous_memory_layout(&datatype->super,1)) {
// fh->f_flags |= OMPIO_CONTIGUOUS_MEMORY;
// }
/**************************************************************************
** In case the data is not contigous in memory, decode it into an iovec **
**************************************************************************/
if (! (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY)) {
ret = fh->f_decode_datatype ((struct mca_io_ompio_file_t *)fh,
datatype,
count,
buf,
&max_data,
&decoded_iov,
&iov_count);
if (OMPI_SUCCESS != ret){
goto exit;
}
}
else {
max_data = count * datatype->super.size;
}
if ( MPI_STATUS_IGNORE != status ) {
status->_ucount = max_data;
}
fh->f_get_num_aggregators ( &dynamic_num_io_procs);
ret = fh->f_set_aggregator_props ((struct mca_io_ompio_file_t *) fh,
dynamic_num_io_procs,
max_data);
if (OMPI_SUCCESS != ret){
goto exit;
}
total_bytes_per_process = (MPI_Aint*)malloc
(fh->f_procs_per_group*sizeof(MPI_Aint));
if (NULL == total_bytes_per_process) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
ret = fh->f_allgather_array (&max_data,
1,
MPI_LONG,
total_bytes_per_process,
1,
MPI_LONG,
fh->f_aggregator_index,
fh->f_procs_in_group,
fh->f_procs_per_group,
fh->f_comm);
if (OMPI_SUCCESS != ret){
goto exit;
}
for (i=0 ; i<fh->f_procs_per_group ; i++) {
total_bytes += total_bytes_per_process[i];
}
if (NULL != total_bytes_per_process) {
free (total_bytes_per_process);
total_bytes_per_process = NULL;
}
/*********************************************************************
*** Generate the File offsets/lengths corresponding to this write ***
********************************************************************/
ret = fh->f_generate_current_file_view ((struct mca_io_ompio_file_t *) fh,
max_data,
&local_iov_array,
&local_count);
if (ret != OMPI_SUCCESS){
goto exit;
}
/* #########################################################*/
/*************************************************************
*** ALLGather the File View information at all processes ***
*************************************************************/
fview_count = (int *) malloc (fh->f_procs_per_group * sizeof (int));
if (NULL == fview_count) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
ret = fh->f_allgather_array (&local_count,
1,
MPI_INT,
fview_count,
1,
MPI_INT,
fh->f_aggregator_index,
fh->f_procs_in_group,
fh->f_procs_per_group,
fh->f_comm);
if (OMPI_SUCCESS != ret){
goto exit;
}
displs = (int*)malloc (fh->f_procs_per_group*sizeof(int));
if (NULL == displs) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
displs[0] = 0;
total_fview_count = fview_count[0];
for (i=1 ; i<fh->f_procs_per_group ; i++) {
total_fview_count += fview_count[i];
displs[i] = displs[i-1] + fview_count[i-1];
}
#if DEBUG_ON
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
for (i=0 ; i<fh->f_procs_per_group ; i++) {
printf ("%d: PROCESS: %d ELEMENTS: %d DISPLS: %d\n",
fh->f_rank,
i,
fview_count[i],
displs[i]);
}
}
#endif
/* allocate the global iovec */
if (0 != total_fview_count) {
global_iov_array = (struct iovec*)malloc (total_fview_count *
sizeof(struct iovec));
if (NULL == global_iov_array) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
}
ret = fh->f_allgatherv_array (local_iov_array,
local_count,
fh->f_iov_type,
global_iov_array,
fview_count,
displs,
fh->f_iov_type,
fh->f_aggregator_index,
fh->f_procs_in_group,
fh->f_procs_per_group,
fh->f_comm);
if (OMPI_SUCCESS != ret){
goto exit;
}
/* sort it */
if (0 != total_fview_count) {
sorted = (int *)malloc (total_fview_count * sizeof(int));
if (NULL == sorted) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
fh->f_sort_iovec (global_iov_array, total_fview_count, sorted);
}
if (NULL != local_iov_array) {
free (local_iov_array);
local_iov_array = NULL;
}
#if DEBUG_ON
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
for (i=0 ; i<total_fview_count ; i++) {
printf("%d: OFFSET: %p LENGTH: %d\n",
fh->f_rank,
global_iov_array[sorted[i]].iov_base,
global_iov_array[sorted[i]].iov_len);
}
}
#endif
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
disp_index = (int *)malloc (fh->f_procs_per_group * sizeof (int));
if (NULL == disp_index) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
blocklen_per_process = (int **)malloc (fh->f_procs_per_group * sizeof (int*));
if (NULL == blocklen_per_process) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
displs_per_process = (MPI_Aint **)malloc (fh->f_procs_per_group * sizeof (MPI_Aint*));
if (NULL == displs_per_process){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
for (i=0;i<fh->f_procs_per_group;i++){
blocklen_per_process[i] = NULL;
displs_per_process[i] = NULL;
}
}
/*
* Calculate how many bytes are read in each cycle
*/
fh->f_get_bytes_per_agg ( (int *) &bytes_per_cycle);
cycles = ceil((double)total_bytes/bytes_per_cycle);
n = 0;
bytes_remaining = 0;
current_index = 0;
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_rexch = MPI_Wtime();
#endif
for (index = 0; index < cycles; index++) {
/* Getting ready for next cycle
Initializing and freeing buffers */
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
if (NULL == sendtype){
sendtype = (ompi_datatype_t **)
malloc (fh->f_procs_per_group * sizeof(ompi_datatype_t *));
if (NULL == sendtype) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
}
for(l=0;l<fh->f_procs_per_group;l++){
disp_index[l] = 1;
if (NULL != blocklen_per_process[l]){
free(blocklen_per_process[l]);
blocklen_per_process[l] = NULL;
}
if (NULL != displs_per_process[l]){
free(displs_per_process[l]);
displs_per_process[l] = NULL;
}
blocklen_per_process[l] = (int *) calloc (1, sizeof(int));
if (NULL == blocklen_per_process[l]) {
opal_output (1, "OUT OF MEMORY for blocklen\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
displs_per_process[l] = (MPI_Aint *) calloc (1, sizeof(MPI_Aint));
if (NULL == displs_per_process[l]){
opal_output (1, "OUT OF MEMORY for displs\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
}
if (NULL != sorted_file_offsets){
free(sorted_file_offsets);
sorted_file_offsets = NULL;
}
if(NULL != file_offsets_for_agg){
free(file_offsets_for_agg);
file_offsets_for_agg = NULL;
}
if (NULL != memory_displacements){
free(memory_displacements);
memory_displacements = NULL;
}
}
if (cycles-1 == index) {
bytes_to_read_in_cycle = total_bytes - bytes_per_cycle*index;
}
else {
bytes_to_read_in_cycle = bytes_per_cycle;
}
#if DEBUG_ON
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
printf ("****%d: CYCLE %d Bytes %d**********\n",
fh->f_rank,
index,
bytes_to_write_in_cycle);
}
#endif
/* Calculate how much data will be contributed in this cycle
by each process*/
bytes_received = 0;
while (bytes_to_read_in_cycle) {
blocks = fview_count[0];
for (j=0 ; j<fh->f_procs_per_group ; j++) {
if (sorted[current_index] < blocks) {
n = j;
break;
}
else {
blocks += fview_count[j+1];
}
}
if (bytes_remaining) {
if (bytes_remaining <= bytes_to_read_in_cycle) {
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
blocklen_per_process[n][disp_index[n] - 1] = bytes_remaining;
displs_per_process[n][disp_index[n] - 1] =
(OPAL_PTRDIFF_TYPE)global_iov_array[sorted[current_index]].iov_base +
(global_iov_array[sorted[current_index]].iov_len - bytes_remaining);
}
if (fh->f_procs_in_group[n] == fh->f_rank) {
bytes_received += bytes_remaining;
}
current_index ++;
bytes_to_read_in_cycle -= bytes_remaining;
bytes_remaining = 0;
if (fh->f_procs_in_group[fh->f_aggregator_index] ==
fh->f_rank) {
blocklen_per_process[n] = (int *) realloc
((void *)blocklen_per_process[n], (disp_index[n]+1)*sizeof(int));
displs_per_process[n] = (MPI_Aint *) realloc
((void *)displs_per_process[n], (disp_index[n]+1)*sizeof(MPI_Aint));
blocklen_per_process[n][disp_index[n]] = 0;
displs_per_process[n][disp_index[n]] = 0;
disp_index[n] += 1;
}
continue;
}
else {
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
blocklen_per_process[n][disp_index[n] - 1] = bytes_to_read_in_cycle;
displs_per_process[n][disp_index[n] - 1] =
(OPAL_PTRDIFF_TYPE)global_iov_array[sorted[current_index]].iov_base +
(global_iov_array[sorted[current_index]].iov_len
- bytes_remaining);
}
if (fh->f_procs_in_group[n] == fh->f_rank) {
bytes_received += bytes_to_read_in_cycle;
}
bytes_remaining -= bytes_to_read_in_cycle;
bytes_to_read_in_cycle = 0;
break;
}
}
else {
if (bytes_to_read_in_cycle <
(MPI_Aint) global_iov_array[sorted[current_index]].iov_len) {
if (fh->f_procs_in_group[fh->f_aggregator_index] ==
fh->f_rank) {
blocklen_per_process[n][disp_index[n] - 1] = bytes_to_read_in_cycle;
displs_per_process[n][disp_index[n] - 1] =
(OPAL_PTRDIFF_TYPE)global_iov_array[sorted[current_index]].iov_base ;
}
if (fh->f_procs_in_group[n] == fh->f_rank) {
bytes_received += bytes_to_read_in_cycle;
}
bytes_remaining = global_iov_array[sorted[current_index]].iov_len -
bytes_to_read_in_cycle;
bytes_to_read_in_cycle = 0;
break;
}
else {
if (fh->f_procs_in_group[fh->f_aggregator_index] ==
fh->f_rank) {
blocklen_per_process[n][disp_index[n] - 1] =
global_iov_array[sorted[current_index]].iov_len;
displs_per_process[n][disp_index[n] - 1] = (OPAL_PTRDIFF_TYPE)
global_iov_array[sorted[current_index]].iov_base;
blocklen_per_process[n] =
(int *) realloc ((void *)blocklen_per_process[n], (disp_index[n]+1)*sizeof(int));
displs_per_process[n] = (MPI_Aint *)realloc
((void *)displs_per_process[n], (disp_index[n]+1)*sizeof(MPI_Aint));
blocklen_per_process[n][disp_index[n]] = 0;
displs_per_process[n][disp_index[n]] = 0;
disp_index[n] += 1;
}
if (fh->f_procs_in_group[n] == fh->f_rank) {
bytes_received +=
global_iov_array[sorted[current_index]].iov_len;
}
bytes_to_read_in_cycle -=
global_iov_array[sorted[current_index]].iov_len;
current_index ++;
continue;
}
}
}
/* Calculate the displacement on where to put the data and allocate
the recieve buffer (global_buf) */
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
entries_per_aggregator=0;
for (i=0;i<fh->f_procs_per_group; i++){
for (j=0;j<disp_index[i];j++){
if (blocklen_per_process[i][j] > 0)
entries_per_aggregator++ ;
}
}
if (entries_per_aggregator > 0){
file_offsets_for_agg = (local_io_array *)
malloc(entries_per_aggregator*sizeof(local_io_array));
if (NULL == file_offsets_for_agg) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
sorted_file_offsets = (int *)
malloc (entries_per_aggregator*sizeof(int));
if (NULL == sorted_file_offsets){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
/*Moving file offsets to an IO array!*/
temp_index = 0;
global_count = 0;
for (i=0;i<fh->f_procs_per_group; i++){
for(j=0;j<disp_index[i];j++){
if (blocklen_per_process[i][j] > 0){
file_offsets_for_agg[temp_index].length =
blocklen_per_process[i][j];
global_count += blocklen_per_process[i][j];
file_offsets_for_agg[temp_index].process_id = i;
file_offsets_for_agg[temp_index].offset =
displs_per_process[i][j];
temp_index++;
}
}
}
}
else{
continue;
}
read_heap_sort (file_offsets_for_agg,
entries_per_aggregator,
sorted_file_offsets);
memory_displacements = (MPI_Aint *) malloc
(entries_per_aggregator * sizeof(MPI_Aint));
memory_displacements[sorted_file_offsets[0]] = 0;
for (i=1; i<entries_per_aggregator; i++){
memory_displacements[sorted_file_offsets[i]] =
memory_displacements[sorted_file_offsets[i-1]] +
file_offsets_for_agg[sorted_file_offsets[i-1]].length;
}
global_buf = (char *) malloc (global_count * sizeof(char));
if (NULL == global_buf){
opal_output(1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
fh->f_io_array = (mca_io_ompio_io_array_t *) malloc
(entries_per_aggregator * sizeof (mca_io_ompio_io_array_t));
if (NULL == fh->f_io_array) {
opal_output(1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
fh->f_num_of_io_entries = 0;
fh->f_io_array[fh->f_num_of_io_entries].offset =
(IOVBASE_TYPE *)(intptr_t)file_offsets_for_agg[sorted_file_offsets[0]].offset;
fh->f_io_array[fh->f_num_of_io_entries].length =
file_offsets_for_agg[sorted_file_offsets[0]].length;
fh->f_io_array[fh->f_num_of_io_entries].memory_address =
global_buf+memory_displacements[sorted_file_offsets[0]];
fh->f_num_of_io_entries++;
for (i=1;i<entries_per_aggregator;i++){
if (file_offsets_for_agg[sorted_file_offsets[i-1]].offset +
file_offsets_for_agg[sorted_file_offsets[i-1]].length ==
file_offsets_for_agg[sorted_file_offsets[i]].offset){
fh->f_io_array[fh->f_num_of_io_entries - 1].length +=
file_offsets_for_agg[sorted_file_offsets[i]].length;
}
else{
fh->f_io_array[fh->f_num_of_io_entries].offset =
(IOVBASE_TYPE *)(intptr_t)file_offsets_for_agg[sorted_file_offsets[i]].offset;
fh->f_io_array[fh->f_num_of_io_entries].length =
file_offsets_for_agg[sorted_file_offsets[i]].length;
fh->f_io_array[fh->f_num_of_io_entries].memory_address =
global_buf+memory_displacements[sorted_file_offsets[i]];
fh->f_num_of_io_entries++;
}
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_read_time = MPI_Wtime();
#endif
if (fh->f_num_of_io_entries) {
if ( 0 > fh->f_fbtl->fbtl_preadv (fh)) {
opal_output (1, "READ FAILED\n");
ret = OMPI_ERROR;
goto exit;
}
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_read_time = MPI_Wtime();
read_time += end_read_time - start_read_time;
#endif
/**********************************************************
******************** DONE READING ************************
*********************************************************/
temp_disp_index = (int *)calloc (1, fh->f_procs_per_group * sizeof (int));
if (NULL == temp_disp_index) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
for (i=0; i<entries_per_aggregator; i++){
temp_index =
file_offsets_for_agg[sorted_file_offsets[i]].process_id;
displs_per_process[temp_index][temp_disp_index[temp_index]] =
memory_displacements[sorted_file_offsets[i]];
if (temp_disp_index[temp_index] < disp_index[temp_index]){
temp_disp_index[temp_index] += 1;
}
else{
printf("temp_disp_index[%d]: %d is greater than disp_index[%d]: %d\n",
temp_index, temp_disp_index[temp_index],
temp_index, disp_index[temp_index]);
}
}
if (NULL != temp_disp_index){
free(temp_disp_index);
temp_disp_index = NULL;
}
send_req = (MPI_Request *)
malloc (fh->f_procs_per_group * sizeof(MPI_Request));
if (NULL == send_req){
opal_output ( 1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_rcomm_time = MPI_Wtime();
#endif
for (i=0;i<fh->f_procs_per_group;i++){
ompi_datatype_create_hindexed(disp_index[i],
blocklen_per_process[i],
displs_per_process[i],
MPI_BYTE,
&sendtype[i]);
ompi_datatype_commit(&sendtype[i]);
ret = MCA_PML_CALL (isend(global_buf,
1,
sendtype[i],
fh->f_procs_in_group[i],
123,
MCA_PML_BASE_SEND_STANDARD,
fh->f_comm,
&send_req[i]));
if(OMPI_SUCCESS != ret){
goto exit;
}
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_rcomm_time = MPI_Wtime();
rcomm_time += end_rcomm_time - start_rcomm_time;
#endif
}
/**********************************************************
********* Scatter the Data from the readers **************
*********************************************************/
if (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY) {
receive_buf = &((char*)buf)[position];
}
else if (bytes_received) {
/* allocate a receive buffer and copy the data that needs
to be received into it in case the data is non-contigous
in memory */
receive_buf = malloc (bytes_received);
if (NULL == receive_buf) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_rcomm_time = MPI_Wtime();
#endif
recv_req = (MPI_Request *) malloc (sizeof (MPI_Request));
if (NULL == recv_req){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
ret = MCA_PML_CALL(irecv(receive_buf,
bytes_received,
MPI_BYTE,
fh->f_procs_in_group[fh->f_aggregator_index],
123,
fh->f_comm,
recv_req));
if (OMPI_SUCCESS != ret){
goto exit;
}
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank){
ret = ompi_request_wait_all (fh->f_procs_per_group,
send_req,
MPI_STATUS_IGNORE);
if (OMPI_SUCCESS != ret){
goto exit;
}
}
ret = ompi_request_wait (recv_req, MPI_STATUS_IGNORE);
if (OMPI_SUCCESS != ret){
goto exit;
}
position += bytes_received;
/* If data is not contigous in memory, copy the data from the
receive buffer into the buffer passed in */
if (!(fh->f_flags & OMPIO_CONTIGUOUS_MEMORY)) {
OPAL_PTRDIFF_TYPE mem_address;
size_t remaining = 0;
size_t temp_position = 0;
remaining = bytes_received;
while (remaining) {
mem_address = (OPAL_PTRDIFF_TYPE)
(decoded_iov[iov_index].iov_base) + current_position;
if (remaining >=
(decoded_iov[iov_index].iov_len - current_position)) {
memcpy ((IOVBASE_TYPE *) mem_address,
receive_buf+temp_position,
decoded_iov[iov_index].iov_len - current_position);
remaining = remaining -
(decoded_iov[iov_index].iov_len - current_position);
temp_position = temp_position +
(decoded_iov[iov_index].iov_len - current_position);
iov_index = iov_index + 1;
current_position = 0;
}
else {
memcpy ((IOVBASE_TYPE *) mem_address,
receive_buf+temp_position,
remaining);
current_position = current_position + remaining;
remaining = 0;
}
}
if (NULL != receive_buf) {
free (receive_buf);
receive_buf = NULL;
}
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_rcomm_time = MPI_Wtime();
rcomm_time += end_rcomm_time - start_rcomm_time;
#endif
if (NULL != recv_req){
free(recv_req);
recv_req = NULL;
}
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank){
fh->f_num_of_io_entries = 0;
if (NULL != fh->f_io_array) {
free (fh->f_io_array);
fh->f_io_array = NULL;
}
if (NULL != global_buf) {
free (global_buf);
global_buf = NULL;
}
for (i = 0; i < fh->f_procs_per_group; i++)
ompi_datatype_destroy(sendtype+i);
if (NULL != sendtype){
free(sendtype);
sendtype=NULL;
}
if (NULL != send_req){
free(send_req);
send_req = NULL;
}
if (NULL != sorted_file_offsets){
free(sorted_file_offsets);
sorted_file_offsets = NULL;
}
if (NULL != file_offsets_for_agg){
free(file_offsets_for_agg);
file_offsets_for_agg = NULL;
}
if (NULL != bytes_per_process){
free(bytes_per_process);
bytes_per_process =NULL;
}
if (NULL != memory_displacements){
free(memory_displacements);
memory_displacements= NULL;
}
}
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_rexch = MPI_Wtime();
read_exch += end_rexch - start_rexch;
nentry.time[0] = read_time;
nentry.time[1] = rcomm_time;
nentry.time[2] = read_exch;
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank)
nentry.aggregator = 1;
else
nentry.aggregator = 0;
nentry.nprocs_for_coll = dynamic_num_io_procs;
if (!fh->f_full_print_queue(READ_PRINT_QUEUE)){
fh->f_register_print_entry(READ_PRINT_QUEUE,
nentry);
}
#endif
exit:
if (NULL != sorted) {
free (sorted);
sorted = NULL;
}
if (NULL != global_iov_array) {
free (global_iov_array);
global_iov_array = NULL;
}
if (NULL != fview_count) {
free (fview_count);
fview_count = NULL;
}
if (NULL != decoded_iov) {
free (decoded_iov);
decoded_iov = NULL;
}
if (NULL != local_iov_array){
free(local_iov_array);
local_iov_array=NULL;
}
if (NULL != displs) {
free (displs);
displs = NULL;
}
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
if (NULL != disp_index){
free(disp_index);
disp_index = NULL;
}
if ( NULL != blocklen_per_process){
for(l=0;l<fh->f_procs_per_group;l++){
if (NULL != blocklen_per_process[l]){
free(blocklen_per_process[l]);
blocklen_per_process[l] = NULL;
}
}
free(blocklen_per_process);
blocklen_per_process = NULL;
}
if (NULL != displs_per_process){
for (l=0; i<fh->f_procs_per_group; l++){
if (NULL != displs_per_process[l]){
free(displs_per_process[l]);
displs_per_process[l] = NULL;
}
}
free(displs_per_process);
displs_per_process = NULL;
}
}
return ret;
}
int ompi_coll_base_alltoall_intra_basic_linear(const void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
void* rbuf, int rcount,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int i, rank, size, err, nreqs, line;
char *psnd, *prcv;
MPI_Aint lb, sndinc, rcvinc;
ompi_request_t **req, **sreq, **rreq;
mca_coll_base_module_t *base_module = (mca_coll_base_module_t*) module;
mca_coll_base_comm_t *data = base_module->base_data;
if (MPI_IN_PLACE == sbuf) {
return mca_coll_base_alltoall_intra_basic_inplace (rbuf, rcount, rdtype,
comm, module);
}
/* Initialize. */
size = ompi_comm_size(comm);
rank = ompi_comm_rank(comm);
OPAL_OUTPUT((ompi_coll_base_framework.framework_output,
"ompi_coll_base_alltoall_intra_basic_linear rank %d", rank));
err = ompi_datatype_get_extent(sdtype, &lb, &sndinc);
if (OMPI_SUCCESS != err) {
return err;
}
sndinc *= scount;
err = ompi_datatype_get_extent(rdtype, &lb, &rcvinc);
if (OMPI_SUCCESS != err) {
return err;
}
rcvinc *= rcount;
/* simple optimization */
psnd = ((char *) sbuf) + (ptrdiff_t)rank * sndinc;
prcv = ((char *) rbuf) + (ptrdiff_t)rank * rcvinc;
err = ompi_datatype_sndrcv(psnd, scount, sdtype, prcv, rcount, rdtype);
if (MPI_SUCCESS != err) {
return err;
}
/* If only one process, we're done. */
if (1 == size) {
return MPI_SUCCESS;
}
/* Initiate all send/recv to/from others. */
req = rreq = coll_base_comm_get_reqs(data, (size - 1) * 2);
prcv = (char *) rbuf;
psnd = (char *) sbuf;
/* Post all receives first -- a simple optimization */
for (nreqs = 0, i = (rank + 1) % size; i != rank;
i = (i + 1) % size, ++rreq, ++nreqs) {
err = MCA_PML_CALL(irecv_init
(prcv + (ptrdiff_t)i * rcvinc, rcount, rdtype, i,
MCA_COLL_BASE_TAG_ALLTOALL, comm, rreq));
if (MPI_SUCCESS != err) { line = __LINE__; goto err_hndl; }
}
/* Now post all sends in reverse order
- We would like to minimize the search time through message queue
when messages actually arrive in the order in which they were posted.
*/
sreq = rreq;
for (i = (rank + size - 1) % size; i != rank;
i = (i + size - 1) % size, ++sreq, ++nreqs) {
err = MCA_PML_CALL(isend_init
(psnd + (ptrdiff_t)i * sndinc, scount, sdtype, i,
MCA_COLL_BASE_TAG_ALLTOALL,
MCA_PML_BASE_SEND_STANDARD, comm, sreq));
if (MPI_SUCCESS != err) { line = __LINE__; goto err_hndl; }
}
/* Start your engines. This will never return an error. */
MCA_PML_CALL(start(nreqs, req));
/* Wait for them all. If there's an error, note that we don't
* care what the error was -- just that there *was* an error. The
* PML will finish all requests, even if one or more of them fail.
* i.e., by the end of this call, all the requests are free-able.
* So free them anyway -- even if there was an error, and return
* the error after we free everything. */
err = ompi_request_wait_all(nreqs, req, MPI_STATUSES_IGNORE);
err_hndl:
if( MPI_SUCCESS != err ) {
OPAL_OUTPUT( (ompi_coll_base_framework.framework_output,"%s:%4d\tError occurred %d, rank %2d",
__FILE__, line, err, rank) );
}
/* Free the reqs in all cases as they are persistent requests */
ompi_coll_base_free_reqs(req, nreqs);
/* All done */
return err;
}
int
mca_fcoll_static_file_read_all (mca_io_ompio_file_t *fh,
void *buf,
int count,
struct ompi_datatype_t *datatype,
ompi_status_public_t *status)
{
int ret = OMPI_SUCCESS, iov_size=0, *bytes_remaining=NULL;
int i, j, l,cycles=0, local_cycles=0, *current_index=NULL;
int index, *disp_index=NULL, *bytes_per_process=NULL, current_position=0;
int **blocklen_per_process=NULL, *iovec_count_per_process=NULL;
int *displs=NULL, *sorted=NULL ,entries_per_aggregator=0;
int *sorted_file_offsets=NULL, temp_index=0, position=0, *temp_disp_index=NULL;
MPI_Aint **displs_per_process=NULL, global_iov_count=0, global_count=0;
MPI_Aint *memory_displacements=NULL;
int bytes_to_read_in_cycle=0;
size_t max_data=0, bytes_per_cycle=0;
uint32_t iov_count=0, iov_index=0;
struct iovec *decoded_iov=NULL, *iov=NULL;
mca_fcoll_static_local_io_array *local_iov_array=NULL, *global_iov_array=NULL;
mca_fcoll_static_local_io_array *file_offsets_for_agg=NULL;
char *global_buf=NULL, *receive_buf=NULL;
int blocklen[3] = {1, 1, 1};
int static_num_io_procs=1;
OPAL_PTRDIFF_TYPE d[3], base;
ompi_datatype_t *types[3];
ompi_datatype_t *io_array_type=MPI_DATATYPE_NULL;
ompi_datatype_t **sendtype = NULL;
MPI_Request *send_req=NULL, recv_req=NULL;
int my_aggregator=-1;
bool recvbuf_is_contiguous=false;
size_t ftype_size;
OPAL_PTRDIFF_TYPE ftype_extent, lb;
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
double read_time = 0.0, start_read_time = 0.0, end_read_time = 0.0;
double rcomm_time = 0.0, start_rcomm_time = 0.0, end_rcomm_time = 0.0;
double read_exch = 0.0, start_rexch = 0.0, end_rexch = 0.0;
mca_common_ompio_print_entry nentry;
#endif
#if DEBUG_ON
MPI_Aint gc_in;
#endif
opal_datatype_type_size ( &datatype->super, &ftype_size );
opal_datatype_get_extent ( &datatype->super, &lb, &ftype_extent );
/**************************************************************************
** 1. In case the data is not contigous in memory, decode it into an iovec
**************************************************************************/
if ( ( ftype_extent == (OPAL_PTRDIFF_TYPE) ftype_size) &&
opal_datatype_is_contiguous_memory_layout(&datatype->super,1) &&
0 == lb ) {
recvbuf_is_contiguous = true;
}
/* In case the data is not contigous in memory, decode it into an iovec */
if (!recvbuf_is_contiguous ) {
fh->f_decode_datatype ( (struct mca_io_ompio_file_t *)fh,
datatype,
count,
buf,
&max_data,
&decoded_iov,
&iov_count);
}
else {
max_data = count * datatype->super.size;
}
if ( MPI_STATUS_IGNORE != status ) {
status->_ucount = max_data;
}
fh->f_get_num_aggregators ( &static_num_io_procs );
fh->f_set_aggregator_props ((struct mca_io_ompio_file_t *) fh,
static_num_io_procs,
max_data);
my_aggregator = fh->f_procs_in_group[fh->f_aggregator_index];
/* printf("max_data %ld\n", max_data); */
ret = fh->f_generate_current_file_view((struct mca_io_ompio_file_t *)fh,
max_data,
&iov,
&iov_size);
if (ret != OMPI_SUCCESS){
goto exit;
}
if ( iov_size > 0 ) {
local_iov_array = (mca_fcoll_static_local_io_array *)malloc (iov_size * sizeof(mca_fcoll_static_local_io_array));
if ( NULL == local_iov_array){
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
for (j=0; j < iov_size; j++){
local_iov_array[j].offset = (OMPI_MPI_OFFSET_TYPE)(intptr_t)
iov[j].iov_base;
local_iov_array[j].length = (size_t)iov[j].iov_len;
local_iov_array[j].process_id = fh->f_rank;
}
}
else {
/* Allocate at least one element to correctly create the derived
data type */
local_iov_array = (mca_fcoll_static_local_io_array *)malloc (sizeof(mca_fcoll_static_local_io_array));
if ( NULL == local_iov_array){
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
local_iov_array[0].offset = (OMPI_MPI_OFFSET_TYPE)(intptr_t) 0;
local_iov_array[0].length = (size_t) 0;
local_iov_array[0].process_id = fh->f_rank;
}
d[0] = (OPAL_PTRDIFF_TYPE)&local_iov_array[0];
d[1] = (OPAL_PTRDIFF_TYPE)&local_iov_array[0].length;
d[2] = (OPAL_PTRDIFF_TYPE)&local_iov_array[0].process_id;
base = d[0];
for (i=0 ; i<3 ; i++) {
d[i] -= base;
}
/* io_array datatype for using in communication*/
types[0] = &ompi_mpi_long.dt;
types[1] = &ompi_mpi_long.dt;
types[2] = &ompi_mpi_int.dt;
ompi_datatype_create_struct (3,
blocklen,
d,
types,
&io_array_type);
ompi_datatype_commit (&io_array_type);
/* #########################################################*/
fh->f_get_bytes_per_agg ( (int*) &bytes_per_cycle);
local_cycles = ceil((double)max_data*fh->f_procs_per_group/bytes_per_cycle);
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_rexch = MPI_Wtime();
#endif
ret = fh->f_comm->c_coll.coll_allreduce (&local_cycles,
&cycles,
1,
MPI_INT,
MPI_MAX,
fh->f_comm,
fh->f_comm->c_coll.coll_allreduce_module);
if (OMPI_SUCCESS != ret){
goto exit;
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_rcomm_time = MPI_Wtime();
rcomm_time += end_rcomm_time - start_rcomm_time;
#endif
if (my_aggregator == fh->f_rank) {
disp_index = (int *) malloc (fh->f_procs_per_group * sizeof(int));
if (NULL == disp_index) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
bytes_per_process = (int *) malloc (fh->f_procs_per_group * sizeof(int ));
if (NULL == bytes_per_process){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
bytes_remaining = (int *) calloc (fh->f_procs_per_group, sizeof(int));
if (NULL == bytes_remaining){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
current_index = (int *) calloc (fh->f_procs_per_group, sizeof(int));
if (NULL == current_index){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
blocklen_per_process = (int **)calloc (fh->f_procs_per_group, sizeof (int*));
if (NULL == blocklen_per_process) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
displs_per_process = (MPI_Aint **)calloc (fh->f_procs_per_group, sizeof (MPI_Aint*));
if (NULL == displs_per_process) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
}
iovec_count_per_process = (int *) calloc (fh->f_procs_per_group, sizeof(int));
if (NULL == iovec_count_per_process){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
displs = (int *) calloc (fh->f_procs_per_group, sizeof(int));
if (NULL == displs){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_rexch = MPI_Wtime();
#endif
ret = fcoll_base_coll_allgather_array (&iov_size,
1,
MPI_INT,
iovec_count_per_process,
1,
MPI_INT,
fh->f_aggregator_index,
fh->f_procs_in_group,
fh->f_procs_per_group,
fh->f_comm);
if( OMPI_SUCCESS != ret){
goto exit;
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_rcomm_time = MPI_Wtime();
rcomm_time += end_rcomm_time - start_rcomm_time;
#endif
if (my_aggregator == fh->f_rank) {
displs[0] = 0;
global_iov_count = iovec_count_per_process[0];
for (i=1 ; i<fh->f_procs_per_group ; i++) {
global_iov_count += iovec_count_per_process[i];
displs[i] = displs[i-1] + iovec_count_per_process[i-1];
}
}
if ( (my_aggregator == fh->f_rank) &&
(global_iov_count > 0 )) {
global_iov_array = (mca_fcoll_static_local_io_array *) malloc (global_iov_count *
sizeof(mca_fcoll_static_local_io_array));
if (NULL == global_iov_array){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_rexch = MPI_Wtime();
#endif
ret = fcoll_base_coll_gatherv_array (local_iov_array,
iov_size,
io_array_type,
global_iov_array,
iovec_count_per_process,
displs,
io_array_type,
fh->f_aggregator_index,
fh->f_procs_in_group,
fh->f_procs_per_group,
fh->f_comm);
if (OMPI_SUCCESS != ret){
fprintf(stderr,"global_iov_array gather error!\n");
goto exit;
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_rcomm_time = MPI_Wtime();
rcomm_time += end_rcomm_time - start_rcomm_time;
#endif
if (NULL != local_iov_array){
free(local_iov_array);
local_iov_array = NULL;
}
if ( ( my_aggregator == fh->f_rank) &&
( global_iov_count > 0 )) {
sorted = (int *)malloc (global_iov_count * sizeof(int));
if (NULL == sorted) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
read_local_heap_sort (global_iov_array, global_iov_count, sorted);
send_req = (MPI_Request *) malloc (fh->f_procs_per_group * sizeof(MPI_Request));
if (NULL == send_req){
opal_output ( 1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
sendtype = (ompi_datatype_t **) malloc (fh->f_procs_per_group * sizeof(ompi_datatype_t *));
if (NULL == sendtype) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
for ( i=0; i<fh->f_procs_per_group; i++ ) {
sendtype[i] = MPI_DATATYPE_NULL;
}
if (NULL == bytes_per_process){
bytes_per_process = (int *) malloc (fh->f_procs_per_group * sizeof(int));
if (NULL == bytes_per_process){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
}
}
#if DEBUG_ON
if (my_aggregator == fh->f_rank) {
for (gc_in=0; gc_in<global_iov_count; gc_in++){
printf("%d: Offset[%ld]: %lld, Length[%ld]: %ld\n",
global_iov_array[sorted[gc_in]].process_id,
gc_in, global_iov_array[sorted[gc_in]].offset,
gc_in, global_iov_array[sorted[gc_in]].length);
}
}
#endif
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_rexch = MPI_Wtime();
#endif
for (index = 0; index < cycles; index++){
if (my_aggregator == fh->f_rank) {
fh->f_num_of_io_entries = 0;
if (NULL != fh->f_io_array) {
free (fh->f_io_array);
fh->f_io_array = NULL;
}
if (NULL != global_buf) {
free (global_buf);
global_buf = NULL;
}
if (NULL != sorted_file_offsets){
free(sorted_file_offsets);
sorted_file_offsets = NULL;
}
if (NULL != file_offsets_for_agg){
free(file_offsets_for_agg);
file_offsets_for_agg = NULL;
}
if (NULL != memory_displacements){
free(memory_displacements);
memory_displacements= NULL;
}
if ( NULL != sendtype ) {
for ( i=0; i<fh->f_procs_per_group; i++ ) {
if ( MPI_DATATYPE_NULL != sendtype[i] ) {
ompi_datatype_destroy (&sendtype[i] );
sendtype[i] = MPI_DATATYPE_NULL;
}
}
}
for(l=0;l<fh->f_procs_per_group;l++){
disp_index[l] = 1;
if (NULL != blocklen_per_process[l]){
free(blocklen_per_process[l]);
blocklen_per_process[l] = NULL;
}
if (NULL != displs_per_process[l]){
free(displs_per_process[l]);
displs_per_process[l] = NULL;
}
blocklen_per_process[l] = (int *) calloc (1, sizeof(int));
if (NULL == blocklen_per_process[l]) {
opal_output (1, "OUT OF MEMORY for blocklen\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
displs_per_process[l] = (MPI_Aint *) calloc (1, sizeof(MPI_Aint));
if (NULL == displs_per_process[l]){
opal_output (1, "OUT OF MEMORY for displs\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
}
}
if (index < local_cycles ) {
if ((index == local_cycles-1) && (max_data % (bytes_per_cycle/fh->f_procs_per_group))) {
bytes_to_read_in_cycle = max_data - position;
}
else if (max_data <= bytes_per_cycle/fh->f_procs_per_group) {
bytes_to_read_in_cycle = max_data;
}
else {
bytes_to_read_in_cycle = bytes_per_cycle/fh->f_procs_per_group;
}
}
else {
bytes_to_read_in_cycle = 0;
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_rexch = MPI_Wtime();
#endif
fcoll_base_coll_gather_array (&bytes_to_read_in_cycle,
1,
MPI_INT,
bytes_per_process,
1,
MPI_INT,
fh->f_aggregator_index,
fh->f_procs_in_group,
fh->f_procs_per_group,
fh->f_comm);
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_rcomm_time = MPI_Wtime();
rcomm_time += end_rcomm_time - start_rcomm_time;
#endif
if (recvbuf_is_contiguous ) {
receive_buf = &((char*)buf)[position];
}
else if (bytes_to_read_in_cycle) {
receive_buf = (char *) malloc (bytes_to_read_in_cycle * sizeof(char));
if ( NULL == receive_buf){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_rcomm_time = MPI_Wtime();
#endif
ret = MCA_PML_CALL(irecv(receive_buf,
bytes_to_read_in_cycle,
MPI_BYTE,
my_aggregator,
123,
fh->f_comm,
&recv_req));
if (OMPI_SUCCESS != ret){
goto exit;
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_rcomm_time = MPI_Wtime();
rcomm_time += end_rcomm_time - start_rcomm_time;
#endif
if (my_aggregator == fh->f_rank) {
for (i=0;i<fh->f_procs_per_group; i++){
while (bytes_per_process[i] > 0){
/*printf("%d: bytes_per_process[%d]: %d, bytes_remaining[%d]: %d\n",
index, i, bytes_per_process[i], i, bytes_remaining[i]);*/
if (read_get_process_id(global_iov_array[sorted[current_index[i]]].process_id,
fh) == i){ /* current id owns this entry!*/
if (bytes_remaining[i]){ /*Remaining bytes in the current entry of
the global offset array*/
if (bytes_remaining[i] <= bytes_per_process[i]){
blocklen_per_process[i][disp_index[i] - 1] = bytes_remaining[i];
displs_per_process[i][disp_index[i] - 1] =
global_iov_array[sorted[current_index[i]]].offset +
(global_iov_array[sorted[current_index[i]]].length
- bytes_remaining[i]);
blocklen_per_process[i] = (int *) realloc
((void *)blocklen_per_process[i], (disp_index[i]+1)*sizeof(int));
displs_per_process[i] = (MPI_Aint *)realloc
((void *)displs_per_process[i], (disp_index[i]+1)*sizeof(MPI_Aint));
bytes_per_process[i] -= bytes_remaining[i];
blocklen_per_process[i][disp_index[i]] = 0;
displs_per_process[i][disp_index[i]] = 0;
disp_index[i] += 1;
bytes_remaining[i] = 0;
/* This entry has been used up, we need to move to the
next entry of this process and make current_index point there*/
current_index[i] = read_find_next_index(i,
current_index[i],
fh,
global_iov_array,
global_iov_count,
sorted);
if (current_index[i] == -1){
break;
}
continue;
}
else{
blocklen_per_process[i][disp_index[i] - 1] = bytes_per_process[i];
displs_per_process[i][disp_index[i] - 1] =
global_iov_array[sorted[current_index[i]]].offset +
(global_iov_array[sorted[current_index[i]]].length
- bytes_remaining[i]);
bytes_remaining[i] -= bytes_per_process[i];
bytes_per_process[i] = 0;
break;
}
}
else{
if (bytes_per_process[i] <
global_iov_array[sorted[current_index[i]]].length){
blocklen_per_process[i][disp_index[i] - 1] =
bytes_per_process[i];
displs_per_process[i][disp_index[i] - 1] =
global_iov_array[sorted[current_index[i]]].offset;
bytes_remaining[i] =
global_iov_array[sorted[current_index[i]]].length -
bytes_per_process[i];
bytes_per_process[i] = 0;
break;
}
else {
blocklen_per_process[i][disp_index[i] - 1] =
global_iov_array[sorted[current_index[i]]].length;
displs_per_process[i][disp_index[i] - 1] =
global_iov_array[sorted[current_index[i]]].offset;
blocklen_per_process[i] =
(int *) realloc ((void *)blocklen_per_process[i], (disp_index[i]+1)*sizeof(int));
displs_per_process[i] = (MPI_Aint *)realloc
((void *)displs_per_process[i], (disp_index[i]+1)*sizeof(MPI_Aint));
blocklen_per_process[i][disp_index[i]] = 0;
displs_per_process[i][disp_index[i]] = 0;
disp_index[i] += 1;
bytes_per_process[i] -=
global_iov_array[sorted[current_index[i]]].length;
current_index[i] = read_find_next_index(i,
current_index[i],
fh,
global_iov_array,
global_iov_count,
sorted);
if (current_index[i] == -1){
break;
}
}
}
}
else{
current_index[i] = read_find_next_index(i,
current_index[i],
fh,
global_iov_array,
global_iov_count,
sorted);
if (current_index[i] == -1){
bytes_per_process[i] = 0; /* no more entries left
to service this request*/
continue;
}
}
}
}
entries_per_aggregator=0;
for (i=0;i<fh->f_procs_per_group;i++){
for (j=0;j<disp_index[i];j++){
if (blocklen_per_process[i][j] > 0){
entries_per_aggregator++;
#if DEBUG_ON
printf("%d sends blocklen[%d]: %d, disp[%d]: %ld to %d\n",
fh->f_procs_in_group[i],j,
blocklen_per_process[i][j],j,
displs_per_process[i][j],
fh->f_rank);
#endif
}
}
}
if (entries_per_aggregator > 0){
file_offsets_for_agg = (mca_fcoll_static_local_io_array *)
malloc(entries_per_aggregator*sizeof(mca_fcoll_static_local_io_array));
if (NULL == file_offsets_for_agg) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
sorted_file_offsets = (int *) malloc (entries_per_aggregator * sizeof(int));
if (NULL == sorted_file_offsets){
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
temp_index=0;
global_count = 0;
for (i=0;i<fh->f_procs_per_group; i++){
for(j=0;j<disp_index[i]; j++){
if (blocklen_per_process[i][j] > 0){
file_offsets_for_agg[temp_index].length =
blocklen_per_process[i][j];
global_count += blocklen_per_process[i][j];
file_offsets_for_agg[temp_index].process_id = i;
file_offsets_for_agg[temp_index].offset =
displs_per_process[i][j];
temp_index++;
}
}
}
}
else{
continue;
}
read_local_heap_sort (file_offsets_for_agg,
entries_per_aggregator,
sorted_file_offsets);
memory_displacements = (MPI_Aint *) malloc
(entries_per_aggregator * sizeof(MPI_Aint));
memory_displacements[sorted_file_offsets[0]] = 0;
for (i=1; i<entries_per_aggregator; i++){
memory_displacements[sorted_file_offsets[i]] =
memory_displacements[sorted_file_offsets[i-1]] +
file_offsets_for_agg[sorted_file_offsets[i-1]].length;
}
global_buf = (char *) malloc (global_count * sizeof(char));
if (NULL == global_buf){
opal_output(1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
#if DEBUG_ON
printf("************Cycle: %d, Aggregator: %d ***************\n",
index+1,fh->f_rank);
for (i=0; i<entries_per_aggregator;i++){
printf("%d: OFFSET: %lld LENGTH: %ld, Mem-offset: %ld, disp_index :%d\n",
file_offsets_for_agg[sorted_file_offsets[i]].process_id,
file_offsets_for_agg[sorted_file_offsets[i]].offset,
file_offsets_for_agg[sorted_file_offsets[i]].length,
memory_displacements[sorted_file_offsets[i]],
disp_index[i]);
}
#endif
fh->f_io_array = (mca_io_ompio_io_array_t *) malloc
(entries_per_aggregator * sizeof (mca_io_ompio_io_array_t));
if (NULL == fh->f_io_array) {
opal_output(1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
fh->f_num_of_io_entries = 0;
fh->f_io_array[0].offset =
(IOVBASE_TYPE *)(intptr_t)file_offsets_for_agg[sorted_file_offsets[0]].offset;
fh->f_io_array[0].length = file_offsets_for_agg[sorted_file_offsets[0]].length;
fh->f_io_array[0].memory_address = global_buf+memory_displacements[sorted_file_offsets[0]];
fh->f_num_of_io_entries++;
for (i=1;i<entries_per_aggregator;i++){
if (file_offsets_for_agg[sorted_file_offsets[i-1]].offset +
file_offsets_for_agg[sorted_file_offsets[i-1]].length ==
file_offsets_for_agg[sorted_file_offsets[i]].offset){
fh->f_io_array[fh->f_num_of_io_entries - 1].length +=
file_offsets_for_agg[sorted_file_offsets[i]].length;
}
else{
fh->f_io_array[fh->f_num_of_io_entries].offset =
(IOVBASE_TYPE *)(intptr_t)file_offsets_for_agg[sorted_file_offsets[i]].offset;
fh->f_io_array[fh->f_num_of_io_entries].length =
file_offsets_for_agg[sorted_file_offsets[i]].length;
fh->f_io_array[fh->f_num_of_io_entries].memory_address =
global_buf+memory_displacements[sorted_file_offsets[i]];
fh->f_num_of_io_entries++;
}
}
#if DEBUG_ON
printf("*************************** %d\n", fh->f_num_of_io_entries);
for (i=0 ; i<fh->f_num_of_io_entries ; i++) {
printf(" ADDRESS: %p OFFSET: %ld LENGTH: %ld\n",
fh->f_io_array[i].memory_address,
(OPAL_PTRDIFF_TYPE)fh->f_io_array[i].offset,
fh->f_io_array[i].length);
}
#endif
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_read_time = MPI_Wtime();
#endif
if (fh->f_num_of_io_entries) {
if ( 0 > fh->f_fbtl->fbtl_preadv (fh)) {
opal_output (1, "READ FAILED\n");
ret = OMPI_ERROR;
goto exit;
}
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_read_time = MPI_Wtime();
read_time += end_read_time - start_read_time;
#endif
#if DEBUG_ON
printf("************Cycle: %d, Aggregator: %d ***************\n",
index+1,fh->f_rank);
if (my_aggregator == fh->f_rank){
for (i=0 ; i<global_count/4 ; i++)
printf (" READ %d \n",((int *)global_buf)[i]);
}
#endif
temp_disp_index = (int *)calloc (1, fh->f_procs_per_group * sizeof (int));
if (NULL == temp_disp_index) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
for (i=0; i<entries_per_aggregator; i++){
temp_index =
file_offsets_for_agg[sorted_file_offsets[i]].process_id;
displs_per_process[temp_index][temp_disp_index[temp_index]] =
memory_displacements[sorted_file_offsets[i]];
if (temp_disp_index[temp_index] < disp_index[temp_index]){
temp_disp_index[temp_index] += 1;
}
else{
printf("temp_disp_index[%d]: %d is greater than disp_index[%d]: %d\n",
temp_index, temp_disp_index[temp_index],
temp_index, disp_index[temp_index]);
}
}
if (NULL != temp_disp_index){
free(temp_disp_index);
temp_disp_index = NULL;
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_rcomm_time = MPI_Wtime();
#endif
for (i=0;i<fh->f_procs_per_group; i++){
send_req[i] = MPI_REQUEST_NULL;
ompi_datatype_create_hindexed(disp_index[i],
blocklen_per_process[i],
displs_per_process[i],
MPI_BYTE,
&sendtype[i]);
ompi_datatype_commit(&sendtype[i]);
ret = MCA_PML_CALL (isend(global_buf,
1,
sendtype[i],
fh->f_procs_in_group[i],
123,
MCA_PML_BASE_SEND_STANDARD,
fh->f_comm,
&send_req[i]));
if(OMPI_SUCCESS != ret){
goto exit;
}
}
ret = ompi_request_wait_all (fh->f_procs_per_group,
send_req,
MPI_STATUS_IGNORE);
if (OMPI_SUCCESS != ret){
goto exit;
}
} /* if ( my_aggregator == fh->f_rank ) */
ret = ompi_request_wait (&recv_req, MPI_STATUS_IGNORE);
if (OMPI_SUCCESS != ret){
goto exit;
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_rcomm_time = MPI_Wtime();
rcomm_time += end_rcomm_time - start_rcomm_time;
#endif
position += bytes_to_read_in_cycle;
if (!recvbuf_is_contiguous) {
OPAL_PTRDIFF_TYPE mem_address;
size_t remaining = 0;
size_t temp_position = 0;
remaining = bytes_to_read_in_cycle;
while (remaining && (iov_count > iov_index)){
mem_address = (OPAL_PTRDIFF_TYPE)
(decoded_iov[iov_index].iov_base) + current_position;
if (remaining >=
(decoded_iov[iov_index].iov_len - current_position)) {
memcpy ((IOVBASE_TYPE *) mem_address,
receive_buf+temp_position,
decoded_iov[iov_index].iov_len - current_position);
remaining = remaining -
(decoded_iov[iov_index].iov_len - current_position);
temp_position = temp_position +
(decoded_iov[iov_index].iov_len - current_position);
iov_index = iov_index + 1;
current_position = 0;
}
else{
memcpy ((IOVBASE_TYPE *) mem_address,
receive_buf+temp_position,
remaining);
current_position = current_position + remaining;
remaining = 0;
}
}
if (NULL != receive_buf) {
free (receive_buf);
receive_buf = NULL;
}
}
}
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
end_rexch = MPI_Wtime();
read_exch += end_rexch - start_rexch;
nentry.time[0] = read_time;
nentry.time[1] = rcomm_time;
nentry.time[2] = read_exch;
if (my_aggregator == fh->f_rank)
nentry.aggregator = 1;
else
nentry.aggregator = 0;
nentry.nprocs_for_coll = static_num_io_procs;
if (!mca_common_ompio_full_print_queue(fh->f_coll_read_time)){
mca_common_ompio_register_print_entry(fh->f_coll_read_time,
nentry);
}
#endif
exit:
if (NULL != decoded_iov){
free(decoded_iov);
decoded_iov = NULL;
}
if (NULL != displs){
free(displs);
displs = NULL;
}
if (NULL != iovec_count_per_process){
free(iovec_count_per_process);
iovec_count_per_process=NULL;
}
if (NULL != local_iov_array){
free(local_iov_array);
local_iov_array=NULL;
}
if (NULL != global_iov_array){
free(global_iov_array);
global_iov_array=NULL;
}
if (my_aggregator == fh->f_rank) {
for(l=0;l<fh->f_procs_per_group;l++){
if (blocklen_per_process) {
free(blocklen_per_process[l]);
}
if (NULL != displs_per_process[l]){
free(displs_per_process[l]);
displs_per_process[l] = NULL;
}
}
}
if (NULL != bytes_per_process){
free(bytes_per_process);
bytes_per_process =NULL;
}
if (NULL != disp_index){
free(disp_index);
disp_index =NULL;
}
if (NULL != displs_per_process){
free(displs_per_process);
displs_per_process = NULL;
}
if(NULL != bytes_remaining){
free(bytes_remaining);
bytes_remaining = NULL;
}
if(NULL != current_index){
free(current_index);
current_index = NULL;
}
if (NULL != blocklen_per_process){
free(blocklen_per_process);
blocklen_per_process =NULL;
}
if (NULL != bytes_remaining){
free(bytes_remaining);
bytes_remaining =NULL;
}
if (NULL != memory_displacements){
free(memory_displacements);
memory_displacements= NULL;
}
if (NULL != file_offsets_for_agg){
free(file_offsets_for_agg);
file_offsets_for_agg = NULL;
}
if (NULL != sorted_file_offsets){
free(sorted_file_offsets);
sorted_file_offsets = NULL;
}
if (NULL != sendtype){
free(sendtype);
sendtype=NULL;
}
if ( !recvbuf_is_contiguous ) {
if (NULL != receive_buf){
free(receive_buf);
receive_buf=NULL;
}
}
if (NULL != global_buf) {
free(global_buf);
global_buf = NULL;
}
if (NULL != sorted) {
free(sorted);
sorted = NULL;
}
if (NULL != send_req){
free(send_req);
send_req = NULL;
}
return ret;
}
int
mca_fcoll_static_file_write_all (mca_io_ompio_file_t *fh,
void *buf,
int count,
struct ompi_datatype_t *datatype,
ompi_status_public_t *status)
{
size_t max_data = 0, bytes_per_cycle=0;
struct iovec *iov=NULL, *decoded_iov=NULL;
uint32_t iov_count=0, iov_index=0;
int i=0,j=0,l=0, temp_index;
int ret=OMPI_SUCCESS, cycles, local_cycles, *bytes_per_process=NULL;
int index, *disp_index=NULL, **blocklen_per_process=NULL;
int *iovec_count_per_process=NULL, *displs=NULL;
size_t total_bytes_written=0;
MPI_Aint **displs_per_process=NULL, *memory_displacements=NULL;
MPI_Aint bytes_to_write_in_cycle=0, global_iov_count=0, global_count=0;
local_io_array *local_iov_array =NULL, *global_iov_array=NULL;
local_io_array *file_offsets_for_agg=NULL;
int *sorted=NULL, *sorted_file_offsets=NULL, temp_pindex, *temp_disp_index=NULL;
char *send_buf=NULL, *global_buf=NULL;
int iov_size=0, current_position=0, *current_index=NULL;
int *bytes_remaining=NULL, entries_per_aggregator=0;
ompi_datatype_t **recvtype = NULL;
MPI_Request *send_req=NULL, *recv_req=NULL;
/* For creating datatype of type io_array */
int blocklen[3] = {1, 1, 1};
int static_num_io_procs=1;
OPAL_PTRDIFF_TYPE d[3], base;
ompi_datatype_t *types[3];
ompi_datatype_t *io_array_type=MPI_DATATYPE_NULL;
/*----------------------------------------------*/
#if TIME_BREAKDOWN
double write_time = 0.0, start_write_time = 0.0, end_write_time = 0.0;
double comm_time = 0.0, start_comm_time = 0.0, end_comm_time = 0.0;
double exch_write = 0.0, start_exch = 0.0, end_exch = 0.0;
print_entry nentry;
#endif
#if DEBUG_ON
MPI_Aint gc_in;
#endif
// if (opal_datatype_is_contiguous_memory_layout(&datatype->super,1)) {
// fh->f_flags |= OMPIO_CONTIGUOUS_MEMORY;
// }
/* In case the data is not contigous in memory, decode it into an iovec */
if (! (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY)) {
ompi_io_ompio_decode_datatype (fh,
datatype,
count,
buf,
&max_data,
&decoded_iov,
&iov_count);
}
else {
max_data = count * datatype->super.size;
}
if ( MPI_STATUS_IGNORE != status ) {
status->_ucount = max_data;
}
mca_io_ompio_get_num_aggregators ( & static_num_io_procs );
ompi_io_ompio_set_aggregator_props (fh,
static_num_io_procs,
max_data);
/* io_array datatype for using in communication*/
types[0] = &ompi_mpi_long.dt;
types[1] = &ompi_mpi_long.dt;
types[2] = &ompi_mpi_int.dt;
d[0] = (OPAL_PTRDIFF_TYPE)&local_iov_array[0];
d[1] = (OPAL_PTRDIFF_TYPE)&local_iov_array[0].length;
d[2] = (OPAL_PTRDIFF_TYPE)&local_iov_array[0].process_id;
base = d[0];
for (i=0 ; i<3 ; i++) {
d[i] -= base;
}
ompi_datatype_create_struct (3,
blocklen,
d,
types,
&io_array_type);
ompi_datatype_commit (&io_array_type);
/* #########################################################*/
ret = ompi_io_ompio_generate_current_file_view(fh,
max_data,
&iov,
&iov_size);
if (ret != OMPI_SUCCESS) {
fprintf(stderr,"Current File View Generation Error\n");
goto exit;
}
if (0 == iov_size) {
iov_size = 1;
}
local_iov_array = (local_io_array *)malloc (iov_size * sizeof(local_io_array));
if ( NULL == local_iov_array) {
fprintf(stderr,"local_iov_array allocation error\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
for (j=0; j < iov_size; j++) {
local_iov_array[j].offset = (OMPI_MPI_OFFSET_TYPE)(intptr_t)
iov[j].iov_base;
local_iov_array[j].length = (size_t)iov[j].iov_len;
local_iov_array[j].process_id = fh->f_rank;
}
mca_io_ompio_get_bytes_per_agg ( (int *) &bytes_per_cycle);
local_cycles = ceil((double)max_data/bytes_per_cycle);
ret = fh->f_comm->c_coll.coll_allreduce (&local_cycles,
&cycles,
1,
MPI_INT,
MPI_MAX,
fh->f_comm,
fh->f_comm->c_coll.coll_allreduce_module);
if (OMPI_SUCCESS != ret) {
fprintf(stderr,"local cycles allreduce!\n");
goto exit;
}
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
disp_index = (int *)malloc (fh->f_procs_per_group * sizeof (int));
if (NULL == disp_index) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
bytes_per_process = (int *) malloc (fh->f_procs_per_group * sizeof(int ));
if (NULL == bytes_per_process) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
bytes_remaining = (int *) malloc (fh->f_procs_per_group * sizeof(int));
if (NULL == bytes_remaining) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
current_index = (int *) malloc (fh->f_procs_per_group * sizeof(int));
if (NULL == current_index) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
blocklen_per_process = (int **)malloc (fh->f_procs_per_group * sizeof (int*));
if (NULL == blocklen_per_process) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
displs_per_process = (MPI_Aint **)
malloc (fh->f_procs_per_group * sizeof (MPI_Aint*));
if (NULL == displs_per_process) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
for(i=0; i<fh->f_procs_per_group; i++) {
current_index[i] = 0;
bytes_remaining[i] =0;
blocklen_per_process[i] = NULL;
displs_per_process[i] = NULL;
}
}
iovec_count_per_process = (int *) malloc (fh->f_procs_per_group * sizeof(int));
if (NULL == iovec_count_per_process) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
displs = (int *) malloc (fh->f_procs_per_group * sizeof(int));
if (NULL == displs) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
ret = ompi_io_ompio_allgather_array (&iov_size,
1,
MPI_INT,
iovec_count_per_process,
1,
MPI_INT,
fh->f_aggregator_index,
fh->f_procs_in_group,
fh->f_procs_per_group,
fh->f_comm);
if( OMPI_SUCCESS != ret) {
fprintf(stderr,"iov size allgatherv array!\n");
goto exit;
}
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
displs[0] = 0;
global_iov_count = iovec_count_per_process[0];
for (i=1 ; i<fh->f_procs_per_group ; i++) {
global_iov_count += iovec_count_per_process[i];
displs[i] = displs[i-1] + iovec_count_per_process[i-1];
}
}
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
global_iov_array = (local_io_array *) malloc (global_iov_count *
sizeof(local_io_array));
if (NULL == global_iov_array) {
opal_output (1, "OUT OF MEMORY\n");
return OMPI_ERR_OUT_OF_RESOURCE;
}
}
ret = ompi_io_ompio_gatherv_array (local_iov_array,
iov_size,
io_array_type,
global_iov_array,
iovec_count_per_process,
displs,
io_array_type,
fh->f_aggregator_index,
fh->f_procs_in_group,
fh->f_procs_per_group,
fh->f_comm);
if (OMPI_SUCCESS != ret) {
fprintf(stderr,"global_iov_array gather error!\n");
goto exit;
}
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
if ( 0 == global_iov_count) {
global_iov_count = 1;
}
sorted = (int *)malloc (global_iov_count * sizeof(int));
if (NULL == sorted) {
opal_output (1, "OUT OF MEMORY\n");
return OMPI_ERR_OUT_OF_RESOURCE;
}
local_heap_sort (global_iov_array, global_iov_count, sorted);
}
#if DEBUG_ON
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
for (gc_in=0; gc_in<global_iov_count; gc_in++) {
printf("%d: Offset[%ld]: %lld, Length[%ld]: %ld\n",
global_iov_array[gc_in].process_id,
gc_in, global_iov_array[gc_in].offset,
gc_in, global_iov_array[gc_in].length);
}
}
#endif
#if TIME_BREAKDOWN
start_exch = MPI_Wtime();
#endif
for (index = 0; index < cycles; index++) {
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
if (NULL == recvtype) {
recvtype = (ompi_datatype_t **)
malloc (fh->f_procs_per_group * sizeof(ompi_datatype_t *));
if (NULL == recvtype) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
}
for(l=0; l<fh->f_procs_per_group; l++) {
disp_index[l] = 1;
if (NULL != blocklen_per_process[l]) {
free(blocklen_per_process[l]);
blocklen_per_process[l] = NULL;
}
if (NULL != displs_per_process[l]) {
free(displs_per_process[l]);
displs_per_process[l] = NULL;
}
blocklen_per_process[l] = (int *) calloc (1, sizeof(int));
if (NULL == blocklen_per_process[l]) {
opal_output (1, "OUT OF MEMORY for blocklen\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
displs_per_process[l] = (MPI_Aint *) calloc (1, sizeof(MPI_Aint));
if (NULL == displs_per_process[l]) {
opal_output (1, "OUT OF MEMORY for displs\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
}
if (NULL != sorted_file_offsets) {
free(sorted_file_offsets);
sorted_file_offsets = NULL;
}
if(NULL != file_offsets_for_agg) {
free(file_offsets_for_agg);
file_offsets_for_agg = NULL;
}
if (NULL != memory_displacements) {
free(memory_displacements);
memory_displacements = NULL;
}
}
if (local_cycles > index) {
if ((index == local_cycles-1) && (max_data % bytes_per_cycle)) {
bytes_to_write_in_cycle = max_data % bytes_per_cycle;
}
else if (max_data <= bytes_per_cycle) {
bytes_to_write_in_cycle = max_data;
}
else {
bytes_to_write_in_cycle = bytes_per_cycle;
}
}
else {
bytes_to_write_in_cycle = 0;
}
#if DEBUG_ON
/* if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {*/
printf ("***%d: CYCLE %d Bytes %ld**********\n",
fh->f_rank,
index,
bytes_to_write_in_cycle);
/* }*/
#endif
/**********************************************************
**Gather the Data from all the processes at the writers **
*********************************************************/
/* gather from each process how many bytes each will be sending */
ompi_io_ompio_gather_array (&bytes_to_write_in_cycle,
1,
MPI_INT,
bytes_per_process,
1,
MPI_INT,
fh->f_aggregator_index,
fh->f_procs_in_group,
fh->f_procs_per_group,
fh->f_comm);
/*
For each aggregator
it needs to get bytes_to_write_in_cycle from each process
in group which adds up to bytes_per_cycle
*/
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
for (i=0; i<fh->f_procs_per_group; i++) {
/* printf("bytes_per_process[%d]: %d\n", i, bytes_per_process[i]);
*/
#if DEBUG_ON
printf ("%d : bytes_per_process : %d\n",
fh->f_procs_in_group[i],
bytes_per_process[i]);
#endif
while (bytes_per_process[i] > 0) {
if (get_process_id(global_iov_array[sorted[current_index[i]]].process_id,
fh) == i) { /* current id owns this entry!*/
/*Add and subtract length and create
blocklength and displs array*/
if (bytes_remaining[i]) {
/*Remaining bytes in the current entry of
the global offset array*/
if (bytes_remaining[i] <= bytes_per_process[i]) {
blocklen_per_process[i][disp_index[i] - 1] = bytes_remaining[i];
displs_per_process[i][disp_index[i] - 1] =
global_iov_array[sorted[current_index[i]]].offset +
(global_iov_array[sorted[current_index[i]]].length
- bytes_remaining[i]);
blocklen_per_process[i] = (int *) realloc
((void *)blocklen_per_process[i], (disp_index[i]+1)*sizeof(int));
displs_per_process[i] = (MPI_Aint *)realloc
((void *)displs_per_process[i], (disp_index[i]+1)*sizeof(MPI_Aint));
bytes_per_process[i] -= bytes_remaining[i];
blocklen_per_process[i][disp_index[i]] = 0;
displs_per_process[i][disp_index[i]] = 0;
bytes_remaining[i] = 0;
disp_index[i] += 1;
/* This entry has been used up, we need to move to the
next entry of this process and make current_index point there*/
current_index[i] = find_next_index(i,
current_index[i],
fh,
global_iov_array,
global_iov_count,
sorted);
if (current_index[i] == -1) {
/* No more entries left, so Its all done! exit!*/
break;
}
continue;
}
else {
blocklen_per_process[i][disp_index[i] - 1] = bytes_per_process[i];
displs_per_process[i][disp_index[i] - 1] =
global_iov_array[sorted[current_index[i]]].offset +
(global_iov_array[sorted[current_index[i]]].length
- bytes_remaining[i]);
bytes_remaining[i] -= bytes_per_process[i];
bytes_per_process[i] = 0;
break;
}
}
else {
if (bytes_per_process[i] <
global_iov_array[sorted[current_index[i]]].length) {
blocklen_per_process[i][disp_index[i] - 1] =
bytes_per_process[i];
displs_per_process[i][disp_index[i] - 1] =
global_iov_array[sorted[current_index[i]]].offset;
bytes_remaining[i] =
global_iov_array[sorted[current_index[i]]].length -
bytes_per_process[i];
bytes_per_process[i] = 0;
break;
}
else {
blocklen_per_process[i][disp_index[i] - 1] =
global_iov_array[sorted[current_index[i]]].length;
displs_per_process[i][disp_index[i] - 1] =
global_iov_array[sorted[current_index[i]]].offset;
blocklen_per_process[i] =
(int *) realloc ((void *)blocklen_per_process[i], (disp_index[i]+1)*sizeof(int));
displs_per_process[i] = (MPI_Aint *)realloc
((void *)displs_per_process[i], (disp_index[i]+1)*sizeof(MPI_Aint));
blocklen_per_process[i][disp_index[i]] = 0;
displs_per_process[i][disp_index[i]] = 0;
disp_index[i] += 1;
bytes_per_process[i] -=
global_iov_array[sorted[current_index[i]]].length;
current_index[i] = find_next_index(i,
current_index[i],
fh,
global_iov_array,
global_iov_count,
sorted);
if (current_index[i] == -1) {
break;
}
}
}
}
else {
current_index[i] = find_next_index(i,
current_index[i],
fh,
global_iov_array,
global_iov_count,
sorted);
if (current_index[i] == -1) {
bytes_per_process[i] = 0; /* no more entries left
to service this request*/
continue;
}
}
}
}
entries_per_aggregator=0;
for (i=0; i<fh->f_procs_per_group; i++) {
for (j=0; j<disp_index[i]; j++) {
if (blocklen_per_process[i][j] > 0) {
entries_per_aggregator++;
#if DEBUG_ON
printf("%d sends blocklen[%d]: %d, disp[%d]: %ld to %d\n",
fh->f_procs_in_group[i],j,
blocklen_per_process[i][j],j,
displs_per_process[i][j],
fh->f_rank);
#endif
}
}
}
if (entries_per_aggregator > 0) {
file_offsets_for_agg = (local_io_array *)
malloc(entries_per_aggregator*sizeof(local_io_array));
if (NULL == file_offsets_for_agg) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
sorted_file_offsets = (int *)
malloc (entries_per_aggregator*sizeof(int));
if (NULL == sorted_file_offsets) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
temp_index = 0;
for (i=0; i<fh->f_procs_per_group; i++) {
for(j=0; j<disp_index[i]; j++) {
if (blocklen_per_process[i][j] > 0) {
file_offsets_for_agg[temp_index].length =
blocklen_per_process[i][j];
file_offsets_for_agg[temp_index].process_id = i;
file_offsets_for_agg[temp_index].offset =
displs_per_process[i][j];
temp_index++;
}
}
}
}
else {
continue;
}
local_heap_sort (file_offsets_for_agg,
entries_per_aggregator,
sorted_file_offsets);
memory_displacements = (MPI_Aint *) malloc
(entries_per_aggregator * sizeof(MPI_Aint));
memory_displacements[sorted_file_offsets[0]] = 0;
for (i=1; i<entries_per_aggregator; i++) {
memory_displacements[sorted_file_offsets[i]] =
memory_displacements[sorted_file_offsets[i-1]] +
file_offsets_for_agg[sorted_file_offsets[i-1]].length;
}
temp_disp_index = (int *)calloc (1, fh->f_procs_per_group * sizeof (int));
if (NULL == temp_disp_index) {
opal_output (1, "OUT OF MEMORY\n");
return OMPI_ERR_OUT_OF_RESOURCE;
}
global_count = 0;
for (i=0; i<entries_per_aggregator; i++) {
temp_pindex =
file_offsets_for_agg[sorted_file_offsets[i]].process_id;
displs_per_process[temp_pindex][temp_disp_index[temp_pindex]] =
memory_displacements[sorted_file_offsets[i]];
if (temp_disp_index[temp_pindex] < disp_index[temp_pindex])
temp_disp_index[temp_pindex] += 1;
else {
printf("temp_disp_index[%d]: %d is greater than disp_index[%d]: %d\n",
temp_pindex, temp_disp_index[temp_pindex],
temp_pindex, disp_index[temp_pindex]);
}
global_count +=
file_offsets_for_agg[sorted_file_offsets[i]].length;
}
if (NULL != temp_disp_index) {
free(temp_disp_index);
temp_disp_index = NULL;
}
#if DEBUG_ON
printf("************Cycle: %d, Aggregator: %d ***************\n",
index+1,fh->f_rank);
for (i=0; i<entries_per_aggregator; i++) {
printf("%d: OFFSET: %lld LENGTH: %ld, Mem-offset: %ld, disp : %d\n",
file_offsets_for_agg[sorted_file_offsets[i]].process_id,
file_offsets_for_agg[sorted_file_offsets[i]].offset,
file_offsets_for_agg[sorted_file_offsets[i]].length,
memory_displacements[sorted_file_offsets[i]],
disp_index[ file_offsets_for_agg[sorted_file_offsets[i]].process_id]);
}
#endif
#if DEBUG_ON
printf("%d: global_count : %ld, bytes_to_write_in_cycle : %ld, procs_per_group: %d\n",
fh->f_rank,
global_count,
bytes_to_write_in_cycle,
fh->f_procs_per_group);
#endif
#if TIME_BREAKDOWN
start_comm_time = MPI_Wtime();
#endif
global_buf = (char *) malloc (global_count);
if (NULL == global_buf) {
opal_output(1, "OUT OF MEMORY");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
recv_req = (MPI_Request *)
malloc (fh->f_procs_per_group * sizeof(MPI_Request));
if (NULL == recv_req) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
for (i=0; i<fh->f_procs_per_group; i++) {
ompi_datatype_create_hindexed(disp_index[i],
blocklen_per_process[i],
displs_per_process[i],
MPI_BYTE,
&recvtype[i]);
ompi_datatype_commit(&recvtype[i]);
ret = MCA_PML_CALL(irecv(global_buf,
1,
recvtype[i],
fh->f_procs_in_group[i],
123,
fh->f_comm,
&recv_req[i]));
if (OMPI_SUCCESS != ret) {
fprintf(stderr,"irecv Error!\n");
goto exit;
}
}
}
if (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY) {
send_buf = &((char*)buf)[total_bytes_written];
}
else if (bytes_to_write_in_cycle) {
/* allocate a send buffer and copy the data that needs
to be sent into it in case the data is non-contigous
in memory */
OPAL_PTRDIFF_TYPE mem_address;
size_t remaining = 0;
size_t temp_position = 0;
send_buf = malloc (bytes_to_write_in_cycle);
if (NULL == send_buf) {
opal_output (1, "OUT OF MEMORY\n");
return OMPI_ERR_OUT_OF_RESOURCE;
}
remaining = bytes_to_write_in_cycle;
while (remaining) {
mem_address = (OPAL_PTRDIFF_TYPE)
(decoded_iov[iov_index].iov_base) + current_position;
if (remaining >=
(decoded_iov[iov_index].iov_len - current_position)) {
memcpy (send_buf+temp_position,
(IOVBASE_TYPE *)mem_address,
decoded_iov[iov_index].iov_len - current_position);
remaining = remaining -
(decoded_iov[iov_index].iov_len - current_position);
temp_position = temp_position +
(decoded_iov[iov_index].iov_len - current_position);
iov_index = iov_index + 1;
current_position = 0;
}
else {
memcpy (send_buf+temp_position,
(IOVBASE_TYPE *)mem_address,
remaining);
current_position = current_position + remaining;
remaining = 0;
}
}
}
total_bytes_written += bytes_to_write_in_cycle;
send_req = (MPI_Request *) malloc (sizeof(MPI_Request));
if (NULL == send_req) {
opal_output (1, "OUT OF MEMORY\n");
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
ret = MCA_PML_CALL(isend(send_buf,
bytes_to_write_in_cycle,
MPI_BYTE,
fh->f_procs_in_group[fh->f_aggregator_index],
123,
MCA_PML_BASE_SEND_STANDARD,
fh->f_comm,
send_req));
if ( OMPI_SUCCESS != ret ) {
fprintf(stderr,"isend error!\n");
goto exit;
}
ret = ompi_request_wait (send_req, MPI_STATUS_IGNORE);
if (OMPI_SUCCESS != ret) {
goto exit;
}
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
ret = ompi_request_wait_all (fh->f_procs_per_group,
recv_req,
MPI_STATUS_IGNORE);
if (OMPI_SUCCESS != ret) {
goto exit;
}
#if DEBUG_ON
printf("************Cycle: %d, Aggregator: %d ***************\n",
index+1,fh->f_rank);
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
for (i=0 ; i<global_count/4 ; i++)
printf (" RECV %d \n",((int *)global_buf)[i]);
}
#endif
}
#if TIME_BREAKDOWN
end_comm_time = MPI_Wtime();
comm_time += end_comm_time - start_comm_time;
#endif
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
fh->f_io_array = (mca_io_ompio_io_array_t *) malloc
(entries_per_aggregator * sizeof (mca_io_ompio_io_array_t));
if (NULL == fh->f_io_array) {
opal_output(1, "OUT OF MEMORY\n");
return OMPI_ERR_OUT_OF_RESOURCE;
}
fh->f_num_of_io_entries = 0;
/*First entry for every aggregator*/
fh->f_io_array[fh->f_num_of_io_entries].offset =
(IOVBASE_TYPE *)(intptr_t)file_offsets_for_agg[sorted_file_offsets[0]].offset;
fh->f_io_array[fh->f_num_of_io_entries].length =
file_offsets_for_agg[sorted_file_offsets[0]].length;
fh->f_io_array[fh->f_num_of_io_entries].memory_address =
global_buf+memory_displacements[sorted_file_offsets[0]];
fh->f_num_of_io_entries++;
for (i=1; i<entries_per_aggregator; i++) {
if (file_offsets_for_agg[sorted_file_offsets[i-1]].offset +
file_offsets_for_agg[sorted_file_offsets[i-1]].length ==
file_offsets_for_agg[sorted_file_offsets[i]].offset) {
fh->f_io_array[fh->f_num_of_io_entries - 1].length +=
file_offsets_for_agg[sorted_file_offsets[i]].length;
}
else {
fh->f_io_array[fh->f_num_of_io_entries].offset =
(IOVBASE_TYPE *)(intptr_t)file_offsets_for_agg[sorted_file_offsets[i]].offset;
fh->f_io_array[fh->f_num_of_io_entries].length =
file_offsets_for_agg[sorted_file_offsets[i]].length;
fh->f_io_array[fh->f_num_of_io_entries].memory_address =
global_buf+memory_displacements[sorted_file_offsets[i]];
fh->f_num_of_io_entries++;
}
}
#if DEBUG_ON
printf("*************************** %d\n", fh->f_num_of_io_entries);
for (i=0 ; i<fh->f_num_of_io_entries ; i++) {
printf(" ADDRESS: %p OFFSET: %ld LENGTH: %ld\n",
fh->f_io_array[i].memory_address,
(OPAL_PTRDIFF_TYPE)fh->f_io_array[i].offset,
fh->f_io_array[i].length);
}
#endif
#if TIME_BREAKDOWN
start_write_time = MPI_Wtime();
#endif
if (fh->f_num_of_io_entries) {
if ( 0 > fh->f_fbtl->fbtl_pwritev (fh)) {
opal_output (1, "WRITE FAILED\n");
ret = OMPI_ERROR;
goto exit;
}
}
#if TIME_BREAKDOWN
end_write_time = MPI_Wtime();
write_time += end_write_time - start_write_time;
#endif
}
if (NULL != send_req) {
free(send_req);
send_req = NULL;
}
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
fh->f_num_of_io_entries = 0;
if (NULL != fh->f_io_array) {
free (fh->f_io_array);
fh->f_io_array = NULL;
}
for (i = 0; i < fh->f_procs_per_group; i++)
ompi_datatype_destroy(recvtype+i);
if (NULL != recvtype) {
free(recvtype);
recvtype=NULL;
}
if (NULL != recv_req) {
free(recv_req);
recv_req = NULL;
}
if (NULL != global_buf) {
free (global_buf);
global_buf = NULL;
}
}
}
#if TIME_BREAKDOWN
end_exch = MPI_Wtime();
exch_write += end_exch - start_exch;
nentry.time[0] = write_time;
nentry.time[1] = comm_time;
nentry.time[2] = exch_write;
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank)
nentry.aggregator = 1;
else
nentry.aggregator = 0;
nentry.nprocs_for_coll = static_num_io_procs;
if (!ompi_io_ompio_full_print_queue(WRITE_PRINT_QUEUE)) {
ompi_io_ompio_register_print_entry(WRITE_PRINT_QUEUE,
nentry);
}
#endif
exit:
if (NULL != decoded_iov) {
free(decoded_iov);
decoded_iov = NULL;
}
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
if (NULL != disp_index) {
free(disp_index);
disp_index = NULL;
}
if (NULL != local_iov_array) {
free(local_iov_array);
local_iov_array = NULL;
}
for(l=0; l<fh->f_procs_per_group; l++) {
if (NULL != blocklen_per_process[l]) {
free(blocklen_per_process[l]);
blocklen_per_process[l] = NULL;
}
if (NULL != displs_per_process[l]) {
free(displs_per_process[l]);
displs_per_process[l] = NULL;
}
}
if (NULL != blocklen_per_process) {
free(blocklen_per_process);
blocklen_per_process = NULL;
}
if (NULL != displs_per_process) {
free(displs_per_process);
displs_per_process = NULL;
}
if(NULL != bytes_remaining) {
free(bytes_remaining);
bytes_remaining = NULL;
}
if(NULL != current_index) {
free(current_index);
current_index = NULL;
}
}
return ret;
}
static int
mca_coll_basic_neighbor_allgather_cart(const void *sbuf, int scount,
struct ompi_datatype_t *sdtype, void *rbuf,
int rcount, struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
const mca_topo_base_comm_cart_2_2_0_t *cart = comm->c_topo->mtc.cart;
const int rank = ompi_comm_rank (comm);
ompi_request_t **reqs, **preqs;
ptrdiff_t lb, extent;
int rc = MPI_SUCCESS, dim, nreqs;
ompi_datatype_get_extent(rdtype, &lb, &extent);
reqs = preqs = coll_base_comm_get_reqs( module->base_data, 4 * cart->ndims );
if( NULL == reqs ) { return OMPI_ERR_OUT_OF_RESOURCE; }
/* The ordering is defined as -1 then +1 in each dimension in
* order of dimension. */
for (dim = 0, nreqs = 0 ; dim < cart->ndims ; ++dim) {
int srank = MPI_PROC_NULL, drank = MPI_PROC_NULL;
if (cart->dims[dim] > 1) {
mca_topo_base_cart_shift (comm, dim, 1, &srank, &drank);
} else if (1 == cart->dims[dim] && cart->periods[dim]) {
srank = drank = rank;
}
if (MPI_PROC_NULL != srank) {
nreqs++;
rc = MCA_PML_CALL(irecv(rbuf, rcount, rdtype, srank,
MCA_COLL_BASE_TAG_ALLGATHER,
comm, preqs++));
if (OMPI_SUCCESS != rc) break;
nreqs++;
/* remove cast from const when the pml layer is updated to take
* a const for the send buffer. */
rc = MCA_PML_CALL(isend((void *) sbuf, scount, sdtype, srank,
MCA_COLL_BASE_TAG_ALLGATHER,
MCA_PML_BASE_SEND_STANDARD,
comm, preqs++));
if (OMPI_SUCCESS != rc) break;
}
rbuf = (char *) rbuf + extent * rcount;
if (MPI_PROC_NULL != drank) {
nreqs++;
rc = MCA_PML_CALL(irecv(rbuf, rcount, rdtype, drank,
MCA_COLL_BASE_TAG_ALLGATHER,
comm, preqs++));
if (OMPI_SUCCESS != rc) break;
nreqs++;
rc = MCA_PML_CALL(isend((void *) sbuf, scount, sdtype, drank,
MCA_COLL_BASE_TAG_ALLGATHER,
MCA_PML_BASE_SEND_STANDARD,
comm, preqs++));
if (OMPI_SUCCESS != rc) break;
}
rbuf = (char *) rbuf + extent * rcount;
}
if (OMPI_SUCCESS != rc) {
ompi_coll_base_free_reqs(reqs, nreqs);
return rc;
}
rc = ompi_request_wait_all (nreqs, reqs, MPI_STATUSES_IGNORE);
if (OMPI_SUCCESS != rc) {
ompi_coll_base_free_reqs(reqs, nreqs);
}
return rc;
}
static int two_phase_exchange_data(mca_io_ompio_file_t *fh,
void *buf, struct iovec *offset_len,
int *send_size, int *start_pos, int *recv_size,
int *count, int *partial_send,
int *recd_from_proc, int contig_access_count,
OMPI_MPI_OFFSET_TYPE min_st_offset,
OMPI_MPI_OFFSET_TYPE fd_size,
OMPI_MPI_OFFSET_TYPE *fd_start,
OMPI_MPI_OFFSET_TYPE *fd_end,
Flatlist_node *flat_buf,
mca_io_ompio_access_array_t *others_req,
int iter, int *buf_idx,
MPI_Aint buftype_extent, int striping_unit,
int *aggregator_list)
{
int i=0, j=0, k=0, tmp=0, nprocs_recv=0, nprocs_send=0;
int ret = OMPI_SUCCESS;
char **recv_buf = NULL;
MPI_Request *requests=NULL;
MPI_Datatype send_type;
#if TIME_BREAKDOWN
start_rcomm_time = MPI_Wtime();
#endif
ret = fh->f_comm->c_coll.coll_alltoall (send_size,
1,
MPI_INT,
recv_size,
1,
MPI_INT,
fh->f_comm,
fh->f_comm->c_coll.coll_alltoall_module);
if ( OMPI_SUCCESS != ret ){
return ret;
}
#if DEBUG
for (i=0; i<fh->f_size; i++){
printf("%d: RS[%d]: %d\n", fh->f_rank,
i,
recv_size[i]);
}
#endif
nprocs_recv = 0;
for (i=0; i < fh->f_size; i++)
if (recv_size[i]) nprocs_recv++;
nprocs_send = 0;
for (i=0; i< fh->f_size; i++)
if (send_size[i]) nprocs_send++;
requests = (MPI_Request *)
malloc((nprocs_send+nprocs_recv+1) * sizeof(MPI_Request));
if (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY) {
j = 0;
for (i=0; i < fh->f_size; i++){
if (recv_size[i]){
ret = MCA_PML_CALL(irecv(((char *) buf)+ buf_idx[i],
recv_size[i],
MPI_BYTE,
i,
fh->f_rank+i+100*iter,
fh->f_comm,
requests+j));
if ( OMPI_SUCCESS != ret ){
return ret;
}
j++;
buf_idx[i] += recv_size[i];
}
}
}
else{
recv_buf = (char **)malloc(fh->f_size * sizeof(char *));
if (NULL == recv_buf){
return OMPI_ERR_OUT_OF_RESOURCE;
}
for (i=0; i < fh->f_size; i++)
if(recv_size[i]) recv_buf[i] =
(char *) malloc (recv_size[i] * sizeof(char));
j = 0;
for(i=0; i<fh->f_size; i++)
if (recv_size[i]) {
ret = MCA_PML_CALL(irecv(recv_buf[i],
recv_size[i],
MPI_BYTE,
i,
fh->f_rank+i+100*iter,
fh->f_comm,
requests+j));
j++;
}
}
j = 0;
for (i = 0; i< fh->f_size; i++){
if (send_size[i]){
if (partial_send[i]){
k = start_pos[i] + count[i] - 1;
tmp = others_req[i].lens[k];
others_req[i].lens[k] = partial_send[i];
}
MPI_Type_hindexed(count[i],
&(others_req[i].lens[start_pos[i]]),
&(others_req[i].mem_ptrs[start_pos[i]]),
MPI_BYTE,
&send_type);
MPI_Type_commit(&send_type);
ret = MCA_PML_CALL(isend(MPI_BOTTOM,
1,
send_type,
i,
fh->f_rank+i+100*iter,
MCA_PML_BASE_SEND_STANDARD,
fh->f_comm,
requests+nprocs_recv+j));
MPI_Type_free(&send_type);
if (partial_send[i]) others_req[i].lens[k] = tmp;
j++;
}
}
if (nprocs_recv) {
ret = ompi_request_wait_all(nprocs_recv,
requests,
MPI_STATUS_IGNORE);
if (! (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY)) {
two_phase_fill_user_buffer(fh, buf, flat_buf,
recv_buf, offset_len,
(unsigned *)recv_size, requests,
recd_from_proc, contig_access_count,
min_st_offset, fd_size, fd_start, fd_end,
buftype_extent, striping_unit, aggregator_list);
}
}
ret = ompi_request_wait_all(nprocs_send,
requests+nprocs_recv,
MPI_STATUS_IGNORE);
if (NULL != requests){
free(requests);
requests = NULL;
}
if (! (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY)){
for (i=0; i< fh->f_size; i++){
if (recv_size[i]){
free(recv_buf[i]);
}
}
free(recv_buf);
}
#if TIME_BREAKDOWN
end_rcomm_time = MPI_Wtime();
rcomm_time += (end_rcomm_time - start_rcomm_time);
#endif
return ret;
}
/**
* All-reduce - subgroup in communicator
*/
OMPI_DECLSPEC int comm_allgather_pml(void *src_buf, void *dest_buf, int count,
ompi_datatype_t *dtype, int my_rank_in_group,
int n_peers, int *ranks_in_comm,ompi_communicator_t *comm)
{
/* local variables */
int rc=OMPI_SUCCESS,msg_cnt;
int pair_rank,exchange,extra_rank, n_extra_nodes,n_extra;
int proc_block,extra_start,extra_end,iovec_len;
int remote_data_start_rank,remote_data_end_rank;
int local_data_start_rank;
netpatterns_pair_exchange_node_t my_exchange_node;
size_t message_extent,current_data_extent,current_data_count;
size_t dt_size;
OPAL_PTRDIFF_TYPE dt_extent;
char *src_buf_current;
char *dest_buf_current;
struct iovec send_iov[2] = {{0,0},{0,0}},
recv_iov[2] = {{0,0},{0,0}};
ompi_request_t *requests[4];
/* get size of data needed - same layout as user data, so that
* we can apply the reudction routines directly on these buffers
*/
rc = ompi_datatype_type_size(dtype, &dt_size);
if( OMPI_SUCCESS != rc ) {
goto Error;
}
rc = ompi_datatype_type_extent(dtype, &dt_extent);
if( OMPI_SUCCESS != rc ) {
goto Error;
}
message_extent = dt_extent*count;
/* place my data in the correct destination buffer */
rc=ompi_datatype_copy_content_same_ddt(dtype,count,
(char *)dest_buf+my_rank_in_group*message_extent,
(char *)src_buf);
if( OMPI_SUCCESS != rc ) {
goto Error;
}
/* 1 process special case */
if(1 == n_peers) {
return OMPI_SUCCESS;
}
/* get my reduction communication pattern */
memset(&my_exchange_node, 0, sizeof(netpatterns_pair_exchange_node_t));
rc = netpatterns_setup_recursive_doubling_tree_node(n_peers,
my_rank_in_group, &my_exchange_node);
if(OMPI_SUCCESS != rc){
return rc;
}
n_extra_nodes=n_peers-my_exchange_node.n_largest_pow_2;
/* get the data from the extra sources */
if(0 < my_exchange_node.n_extra_sources) {
if ( EXCHANGE_NODE == my_exchange_node.node_type ) {
/*
** Receive data from extra node
*/
extra_rank=my_exchange_node.rank_extra_source;
/* receive the data into the correct location - will use 2
* messages in the recursive doubling phase */
dest_buf_current=(char *)dest_buf+message_extent*extra_rank;
rc=MCA_PML_CALL(recv(dest_buf_current,
count,dtype,ranks_in_comm[extra_rank],
-OMPI_COMMON_TAG_ALLREDUCE,
comm, MPI_STATUSES_IGNORE));
if( 0 > rc ) {
goto Error;
}
} else {
/*
** Send data to "partner" node
*/
extra_rank=my_exchange_node.rank_extra_source;
src_buf_current=(char *)src_buf;
rc=MCA_PML_CALL(send(src_buf_current,
count,dtype,ranks_in_comm[extra_rank],
-OMPI_COMMON_TAG_ALLREDUCE,
MCA_PML_BASE_SEND_STANDARD,
comm));
if( 0 > rc ) {
goto Error;
}
}
}
current_data_extent=message_extent;
current_data_count=count;
src_buf_current=(char *)dest_buf+my_rank_in_group*message_extent;
proc_block=1;
local_data_start_rank=my_rank_in_group;
/* loop over data exchanges */
for(exchange=0 ; exchange < my_exchange_node.n_exchanges ; exchange++) {
/* is the remote data read */
pair_rank=my_exchange_node.rank_exchanges[exchange];
msg_cnt=0;
/*
* Power of 2 data segment
*/
/* post non-blocking receive */
if(pair_rank > my_rank_in_group ){
recv_iov[0].iov_base=src_buf_current+current_data_extent;
recv_iov[0].iov_len=current_data_extent;
iovec_len=1;
remote_data_start_rank=local_data_start_rank+proc_block;
remote_data_end_rank=remote_data_start_rank+proc_block-1;
} else {
recv_iov[0].iov_base=src_buf_current-current_data_extent;
recv_iov[0].iov_len=current_data_extent;
iovec_len=1;
remote_data_start_rank=local_data_start_rank-proc_block;
remote_data_end_rank=remote_data_start_rank+proc_block-1;
}
/* the data from the non power of 2 ranks */
if(remote_data_start_rank<n_extra_nodes) {
/* figure out how much data is at the remote rank */
/* last rank with data */
extra_start=remote_data_start_rank;
extra_end=remote_data_end_rank;
if(extra_end >= n_extra_nodes ) {
/* if last rank exceeds the ranks with extra data,
* adjust this.
*/
extra_end=n_extra_nodes-1;
}
/* get the number of ranks whos data is to be grabbed */
n_extra=extra_end-extra_start+1;
recv_iov[1].iov_base=(char *)dest_buf+
(extra_start+my_exchange_node.n_largest_pow_2)*message_extent;
recv_iov[1].iov_len=n_extra*count;
iovec_len=2;
}
rc=MCA_PML_CALL(irecv(recv_iov[0].iov_base,
current_data_count,dtype,ranks_in_comm[pair_rank],
-OMPI_COMMON_TAG_ALLREDUCE,
comm,&(requests[msg_cnt])));
if( 0 > rc ) {
goto Error;
}
msg_cnt++;
if(iovec_len > 1 ) {
rc=MCA_PML_CALL(irecv(recv_iov[1].iov_base,
recv_iov[1].iov_len,dtype,ranks_in_comm[pair_rank],
-OMPI_COMMON_TAG_ALLREDUCE,
comm,&(requests[msg_cnt])));
if( 0 > rc ) {
goto Error;
}
msg_cnt++;
}
/* post non-blocking send */
send_iov[0].iov_base=src_buf_current;
send_iov[0].iov_len=current_data_extent;
iovec_len=1;
/* the data from the non power of 2 ranks */
if(local_data_start_rank<n_extra_nodes) {
/* figure out how much data is at the remote rank */
/* last rank with data */
extra_start=local_data_start_rank;
extra_end=extra_start+proc_block-1;
if(extra_end >= n_extra_nodes ) {
/* if last rank exceeds the ranks with extra data,
* adjust this.
*/
extra_end=n_extra_nodes-1;
}
/* get the number of ranks whos data is to be grabbed */
n_extra=extra_end-extra_start+1;
send_iov[1].iov_base=(char *)dest_buf+
(extra_start+my_exchange_node.n_largest_pow_2)*message_extent;
send_iov[1].iov_len=n_extra*count;
iovec_len=2;
}
rc=MCA_PML_CALL(isend(send_iov[0].iov_base,
current_data_count,dtype,ranks_in_comm[pair_rank],
-OMPI_COMMON_TAG_ALLREDUCE,MCA_PML_BASE_SEND_STANDARD,
comm,&(requests[msg_cnt])));
if( 0 > rc ) {
goto Error;
}
msg_cnt++;
if( iovec_len > 1 ) {
rc=MCA_PML_CALL(isend(send_iov[1].iov_base,
send_iov[1].iov_len,dtype,ranks_in_comm[pair_rank],
-OMPI_COMMON_TAG_ALLREDUCE,MCA_PML_BASE_SEND_STANDARD,
comm,&(requests[msg_cnt])));
if( 0 > rc ) {
goto Error;
}
msg_cnt++;
}
/* prepare the source buffer for the next iteration */
if(pair_rank < my_rank_in_group ){
src_buf_current-=current_data_extent;
local_data_start_rank-=proc_block;
}
proc_block*=2;
current_data_extent*=2;
current_data_count*=2;
/* wait on send and receive completion */
ompi_request_wait_all(msg_cnt,requests,MPI_STATUSES_IGNORE);
}
/* copy data in from the "extra" source, if need be */
if(0 < my_exchange_node.n_extra_sources) {
if ( EXTRA_NODE == my_exchange_node.node_type ) {
/*
** receive the data
** */
extra_rank=my_exchange_node.rank_extra_source;
rc=MCA_PML_CALL(recv(dest_buf,
count*n_peers,dtype,ranks_in_comm[extra_rank],
-OMPI_COMMON_TAG_ALLREDUCE,
comm,MPI_STATUSES_IGNORE));
if(0 > rc ) {
goto Error;
}
} else {
/* send the data to the pair-rank outside of the power of 2 set
** of ranks
*/
extra_rank=my_exchange_node.rank_extra_source;
rc=MCA_PML_CALL(send(dest_buf,
count*n_peers,dtype,ranks_in_comm[extra_rank],
-OMPI_COMMON_TAG_ALLREDUCE,
MCA_PML_BASE_SEND_STANDARD,
comm));
if( 0 > rc ) {
goto Error;
}
}
}
netpatterns_cleanup_recursive_doubling_tree_node(&my_exchange_node);
/* return */
return OMPI_SUCCESS;
Error:
return rc;
}