int
mca_fcoll_dynamic_file_write_all (mca_io_ompio_file_t *fh,
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;
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 recv_req_count=0;
#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 (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 = ompi_io_ompio_decode_datatype (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;
}
mca_io_ompio_get_num_aggregators ( &dynamic_num_io_procs );
ret = ompi_io_ompio_set_aggregator_props (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 = ompi_io_ompio_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 = ompi_io_ompio_generate_current_file_view(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
/*************************************************************
*** 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 = ompi_io_ompio_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
printf("total_fview_count : %d\n", total_fview_count);
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 = ompi_io_ompio_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;
}
ompi_io_ompio_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 (fh->f_procs_in_group[fh->f_aggregator_index] == 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
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;
}
}
mca_io_ompio_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 TIME_BREAKDOWN
start_exch = 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 == 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 (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 (fh->f_procs_in_group[fh->f_aggregator_index] == 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 **
*********************************************************/
/* Calculate how much data will be contributed in this cycle
by each process*/
bytes_sent = 0;
#if DEBUG_ON
printf("bytes_to_write_in_cycle: %ld, cycle : %d\n", bytes_to_write_in_cycle,
index);
#endif
/* The blocklen and displs calculation only done at aggregators!*/
while (bytes_to_write_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_write_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_sent += bytes_remaining;
}
current_index ++;
bytes_to_write_in_cycle -= bytes_remaining;
bytes_remaining = 0;
if (fh->f_procs_in_group[fh->f_aggregator_index] ==
fh->f_rank) {
/* 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;
}
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_write_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_sent += bytes_to_write_in_cycle;
}
bytes_remaining -= bytes_to_write_in_cycle;
bytes_to_write_in_cycle = 0;
break;
}
}
else {
if (bytes_to_write_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_write_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_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 {
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;
/*realloc for next blocklength
and assign this displacement and check for next displs as
the total length of this entry has been consumed!*/
}
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;
}
}
}
/* 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 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 = (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;
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");
return OMPI_ERR_OUT_OF_RESOURCE;
}
/*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 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_count = 0;
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]);
opal_datatype_type_size(&recvtype[i]->super,
&datatype_size);
if (datatype_size){
recv_req = (MPI_Request *)realloc
((void *)recv_req, (recv_req_count + 1)*sizeof(MPI_Request));
ret = MCA_PML_CALL(irecv(global_buf,
1,
recvtype[i],
fh->f_procs_in_group[i],
123,
fh->f_comm,
&recv_req[recv_req_count]));
recv_req_count++;
if (OMPI_SUCCESS != ret){
goto exit;
}
}
}
}
if (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY) {
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 */
OPAL_PTRDIFF_TYPE 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");
return OMPI_ERR_OUT_OF_RESOURCE;
}
remaining = bytes_sent;
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_sent;
/* Gather the sendbuf from each process in appropritate locations in
aggregators*/
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;
}
if (bytes_sent){
ret = MCA_PML_CALL(isend(send_buf,
bytes_sent,
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 ){
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 (recv_req_count,
recv_req,
MPI_STATUS_IGNORE);
if (OMPI_SUCCESS != ret){
goto exit;
}
}
#if DEBUG_ON
if (fh->f_procs_in_group[fh->f_aggregator_index] == 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 (! (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY)) {
if (NULL != send_buf) {
free (send_buf);
send_buf = NULL;
}
}
#if TIME_BREAKDOWN
end_comm_time = MPI_Wtime();
comm_time += (end_comm_time - start_comm_time);
#endif
/**********************************************************
**************** DONE GATHERING OF DATA ******************
*********************************************************/
/**********************************************************
******* Create the io array, and pass it to fbtl *********
*********************************************************/
if (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
#if TIME_BREAKDOWN
start_write_time = MPI_Wtime();
#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");
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 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,
(OPAL_PTRDIFF_TYPE)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 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 = dynamic_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 (fh->f_procs_in_group[fh->f_aggregator_index] == fh->f_rank) {
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[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 != 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 != send_req){
free(send_req);
send_req = NULL;
}
return OMPI_SUCCESS;
}
int
mca_fcoll_two_phase_file_write_all (mca_io_ompio_file_t *fh,
void *buf,
int count,
struct ompi_datatype_t *datatype,
ompi_status_public_t *status)
{
int i, j,interleave_count=0, striping_unit=0;
uint32_t iov_count=0,ti;
struct iovec *decoded_iov=NULL, *temp_iov=NULL;
size_t max_data = 0, total_bytes = 0;
int domain_size=0, *count_my_req_per_proc=NULL, count_my_req_procs;
int count_other_req_procs, ret=OMPI_SUCCESS;
size_t *buf_indices=NULL;
int local_count = 0, local_size=0,*aggregator_list = NULL;
struct iovec *iov = NULL;
OMPI_MPI_OFFSET_TYPE start_offset, end_offset, fd_size;
OMPI_MPI_OFFSET_TYPE *start_offsets=NULL, *end_offsets=NULL;
OMPI_MPI_OFFSET_TYPE *fd_start=NULL, *fd_end=NULL, min_st_offset;
Flatlist_node *flat_buf=NULL;
mca_io_ompio_access_array_t *my_req=NULL, *others_req=NULL;
MPI_Aint send_buf_addr;
#if TIME_BREAKDOWN
print_entry nentry;
#endif
if (opal_datatype_is_contiguous_memory_layout(&datatype->super,1)) {
fh->f_flags = fh->f_flags | OMPIO_CONTIGUOUS_MEMORY;
}
if (! (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY)) {
ret = ompi_io_ompio_decode_datatype (fh,
datatype,
count,
buf,
&max_data,
&temp_iov,
&iov_count);
if (OMPI_SUCCESS != ret ){
goto exit;
}
send_buf_addr = (OPAL_PTRDIFF_TYPE)buf;
decoded_iov = (struct iovec *)malloc
(iov_count * sizeof(struct iovec));
for (ti = 0; ti < iov_count; ti ++){
decoded_iov[ti].iov_base = (IOVBASE_TYPE *)(
(OPAL_PTRDIFF_TYPE)temp_iov[ti].iov_base -
send_buf_addr);
decoded_iov[ti].iov_len =
temp_iov[ti].iov_len ;
#if DEBUG_ON
printf("d_offset[%d]: %ld, d_len[%d]: %ld\n",
ti, (OPAL_PTRDIFF_TYPE)decoded_iov[ti].iov_base,
ti, decoded_iov[ti].iov_len);
#endif
}
}
else{
max_data = count * datatype->super.size;
}
if ( MPI_STATUS_IGNORE != status ) {
status->_ucount = max_data;
}
if(-1 == mca_fcoll_two_phase_num_io_procs){
ret = ompi_io_ompio_set_aggregator_props (fh,
mca_fcoll_two_phase_num_io_procs,
max_data);
if ( OMPI_SUCCESS != ret){
return ret;
}
mca_fcoll_two_phase_num_io_procs =
ceil((float)fh->f_size/fh->f_procs_per_group);
}
if (mca_fcoll_two_phase_num_io_procs > fh->f_size){
mca_fcoll_two_phase_num_io_procs = fh->f_size;
}
#if DEBUG_ON
printf("Number of aggregators : %ld\n", mca_fcoll_two_phase_num_io_procs);
#endif
aggregator_list = (int *) malloc (mca_fcoll_two_phase_num_io_procs *
sizeof(int));
if ( NULL == aggregator_list ) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
for (i =0; i< mca_fcoll_two_phase_num_io_procs; i++){
aggregator_list[i] = i;
}
ret = ompi_io_ompio_generate_current_file_view (fh,
max_data,
&iov,
&local_count);
if ( OMPI_SUCCESS != ret ){
goto exit;
}
ret = fh->f_comm->c_coll.coll_allreduce (&max_data,
&total_bytes,
1,
MPI_DOUBLE,
MPI_SUM,
fh->f_comm,
fh->f_comm->c_coll.coll_allreduce_module);
if ( OMPI_SUCCESS != ret ) {
goto exit;
}
if (!(fh->f_flags & OMPIO_CONTIGUOUS_MEMORY)) {
/* This datastructre translates between OMPIO->ROMIO its a little hacky!*/
/* But helps to re-use romio's code for handling non-contiguous file-type*/
flat_buf = (Flatlist_node *)malloc(sizeof(Flatlist_node));
if ( NULL == flat_buf ){
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
flat_buf->type = datatype;
flat_buf->next = NULL;
flat_buf->count = 0;
local_size = iov_count/count;
flat_buf->indices =
(OMPI_MPI_OFFSET_TYPE *)malloc(local_size *
sizeof(OMPI_MPI_OFFSET_TYPE));
if ( NULL == flat_buf->indices ){
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
flat_buf->blocklens =
(OMPI_MPI_OFFSET_TYPE *)malloc(local_size *
sizeof(OMPI_MPI_OFFSET_TYPE));
if ( NULL == flat_buf->blocklens ){
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
flat_buf->count = local_size;
i=0;j=0;
while(j < local_size){
flat_buf->indices[j] = (OMPI_MPI_OFFSET_TYPE)(intptr_t)decoded_iov[i].iov_base;
flat_buf->blocklens[j] = decoded_iov[i].iov_len;
if(i < (int)iov_count)
i+=1;
j+=1;
}
#if DEBUG_ON
printf("flat_buf_count : %d\n", flat_buf->count);
for(i=0;i<flat_buf->count;i++){
printf("%d: blocklen[%d] : %lld, indices[%d]: %lld \n",
fh->f_rank, i, flat_buf->blocklens[i], i ,flat_buf->indices[i]);
}
#endif
}
#if DEBUG_ON
printf("%d: fcoll:two_phase:write_all->total_bytes:%ld, local_count: %d\n",
fh->f_rank,total_bytes, local_count);
for (i=0 ; i<local_count ; i++) {
printf("%d: fcoll:two_phase:write_all:OFFSET:%ld,LENGTH:%ld\n",
fh->f_rank,
(size_t)iov[i].iov_base,
(size_t)iov[i].iov_len);
}
#endif
start_offset = (OMPI_MPI_OFFSET_TYPE)(intptr_t)iov[0].iov_base;
end_offset = (OMPI_MPI_OFFSET_TYPE)(intptr_t)iov[local_count-1].iov_base +
(OMPI_MPI_OFFSET_TYPE)iov[local_count-1].iov_len - 1;
#if DEBUG_ON
printf("%d: fcoll:two_phase:write_all:START OFFSET:%ld,END OFFSET:%ld\n",
fh->f_rank,
(size_t)start_offset,
(size_t)end_offset);
#endif
start_offsets = (OMPI_MPI_OFFSET_TYPE *)malloc
(fh->f_size*sizeof(OMPI_MPI_OFFSET_TYPE));
if ( NULL == start_offsets ){
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
end_offsets = (OMPI_MPI_OFFSET_TYPE *)malloc
(fh->f_size*sizeof(OMPI_MPI_OFFSET_TYPE));
if ( NULL == end_offsets ){
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
ret = fh->f_comm->c_coll.coll_allgather(&start_offset,
1,
MPI_LONG,
start_offsets,
1,
MPI_LONG,
fh->f_comm,
fh->f_comm->c_coll.coll_allgather_module);
if ( OMPI_SUCCESS != ret ){
goto exit;
}
ret = fh->f_comm->c_coll.coll_allgather(&end_offset,
1,
MPI_LONG,
end_offsets,
1,
MPI_LONG,
fh->f_comm,
fh->f_comm->c_coll.coll_allgather_module);
if ( OMPI_SUCCESS != ret ){
goto exit;
}
#if DEBUG_ON
for (i=0;i<fh->f_size;i++){
printf("%d: fcoll:two_phase:write_all:start[%d]:%ld,end[%d]:%ld\n",
fh->f_rank,i,
(size_t)start_offsets[i],i,
(size_t)end_offsets[i]);
}
#endif
for (i=1; i<fh->f_size; i++){
if ((start_offsets[i] < end_offsets[i-1]) &&
(start_offsets[i] <= end_offsets[i])){
interleave_count++;
}
}
#if DEBUG_ON
printf("%d: fcoll:two_phase:write_all:interleave_count:%d\n",
fh->f_rank,interleave_count);
#endif
ret = mca_fcoll_two_phase_domain_partition(fh,
start_offsets,
end_offsets,
&min_st_offset,
&fd_start,
&fd_end,
domain_size,
&fd_size,
striping_unit,
mca_fcoll_two_phase_num_io_procs);
if ( OMPI_SUCCESS != ret ){
goto exit;
}
#if DEBUG_ON
for (i=0;i<mca_fcoll_two_phase_num_io_procs;i++){
printf("fd_start[%d] : %lld, fd_end[%d] : %lld, local_count: %d\n",
i, fd_start[i], i, fd_end[i], local_count);
}
#endif
ret = mca_fcoll_two_phase_calc_my_requests (fh,
iov,
local_count,
min_st_offset,
fd_start,
fd_end,
fd_size,
&count_my_req_procs,
&count_my_req_per_proc,
&my_req,
&buf_indices,
striping_unit,
mca_fcoll_two_phase_num_io_procs,
aggregator_list);
if ( OMPI_SUCCESS != ret ){
goto exit;
}
ret = mca_fcoll_two_phase_calc_others_requests(fh,
count_my_req_procs,
count_my_req_per_proc,
my_req,
&count_other_req_procs,
&others_req);
if (OMPI_SUCCESS != ret ){
goto exit;
}
#if DEBUG_ON
printf("count_other_req_procs : %d\n", count_other_req_procs);
#endif
#if TIME_BREAKDOWN
start_exch = MPI_Wtime();
#endif
ret = two_phase_exch_and_write(fh,
buf,
datatype,
others_req,
iov,
local_count,
min_st_offset,
fd_size,
fd_start,
fd_end,
flat_buf,
buf_indices,
striping_unit,
aggregator_list);
if (OMPI_SUCCESS != ret){
goto exit;
}
#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 (is_aggregator(fh->f_rank,
mca_fcoll_two_phase_num_io_procs,
aggregator_list)){
nentry.aggregator = 1;
}
else{
nentry.aggregator = 0;
}
nentry.nprocs_for_coll = mca_fcoll_two_phase_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 (flat_buf != NULL) {
if (flat_buf->blocklens != NULL) {
free (flat_buf->blocklens);
}
if (flat_buf->indices != NULL) {
free (flat_buf->indices);
}
free (flat_buf);
}
if (start_offsets != NULL) {
free(start_offsets);
}
if (end_offsets != NULL){
free(end_offsets);
}
if (aggregator_list != NULL){
free(aggregator_list);
}
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;
}
