int mca_fcoll_two_phase_domain_partition (mca_io_ompio_file_t *fh,
OMPI_MPI_OFFSET_TYPE *start_offsets,
OMPI_MPI_OFFSET_TYPE *end_offsets,
OMPI_MPI_OFFSET_TYPE *min_st_offset_ptr,
OMPI_MPI_OFFSET_TYPE **fd_st_ptr,
OMPI_MPI_OFFSET_TYPE **fd_end_ptr,
int min_fd_size,
OMPI_MPI_OFFSET_TYPE *fd_size_ptr,
int striping_unit,
int nprocs_for_coll){
OMPI_MPI_OFFSET_TYPE min_st_offset, max_end_offset, *fd_start=NULL, *fd_end=NULL, fd_size;
int i;
min_st_offset = start_offsets[0];
max_end_offset = end_offsets[0];
for (i=0; i< fh->f_size; i++){
min_st_offset = OMPIO_MIN(min_st_offset, start_offsets[i]);
max_end_offset = OMPIO_MAX(max_end_offset, end_offsets[i]);
}
fd_size = ((max_end_offset - min_st_offset + 1) + nprocs_for_coll - 1)/nprocs_for_coll;
if (fd_size < min_fd_size)
fd_size = min_fd_size;
*fd_st_ptr = (OMPI_MPI_OFFSET_TYPE *)
malloc(nprocs_for_coll*sizeof(OMPI_MPI_OFFSET_TYPE));
if ( NULL == *fd_st_ptr ) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
*fd_end_ptr = (OMPI_MPI_OFFSET_TYPE *)
malloc(nprocs_for_coll*sizeof(OMPI_MPI_OFFSET_TYPE));
if ( NULL == *fd_end_ptr ) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
fd_start = *fd_st_ptr;
fd_end = *fd_end_ptr;
if (striping_unit > 0){
/* Lock Boundary based domain partitioning */
int rem_front, rem_back;
OMPI_MPI_OFFSET_TYPE end_off;
fd_start[0] = min_st_offset;
end_off = fd_start[0] + fd_size;
rem_front = end_off % striping_unit;
rem_back = striping_unit - rem_front;
if (rem_front < rem_back)
end_off -= rem_front;
else
end_off += rem_back;
fd_end[0] = end_off - 1;
/* align fd_end[i] to the nearest file lock boundary */
for (i=1; i<nprocs_for_coll; i++) {
fd_start[i] = fd_end[i-1] + 1;
end_off = min_st_offset + fd_size * (i+1);
rem_front = end_off % striping_unit;
rem_back = striping_unit - rem_front;
if (rem_front < rem_back)
end_off -= rem_front;
else
end_off += rem_back;
fd_end[i] = end_off - 1;
}
fd_end[nprocs_for_coll-1] = max_end_offset;
}
else{
fd_start[0] = min_st_offset;
fd_end[0] = min_st_offset + fd_size - 1;
for (i=1; i<nprocs_for_coll; i++) {
fd_start[i] = fd_end[i-1] + 1;
fd_end[i] = fd_start[i] + fd_size - 1;
}
}
for (i=0; i<nprocs_for_coll; i++) {
if (fd_start[i] > max_end_offset)
fd_start[i] = fd_end[i] = -1;
if (fd_end[i] > max_end_offset)
fd_end[i] = max_end_offset;
}
*fd_size_ptr = fd_size;
*min_st_offset_ptr = min_st_offset;
return OMPI_SUCCESS;
}
static int two_phase_read_and_exch(mca_io_ompio_file_t *fh,
void *buf,
MPI_Datatype datatype,
mca_io_ompio_access_array_t *others_req,
struct iovec *offset_len,
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,
size_t *buf_idx, int striping_unit,
int *aggregator_list){
int ret=OMPI_SUCCESS, i = 0, j = 0, ntimes = 0, max_ntimes = 0;
int m = 0;
int *curr_offlen_ptr=NULL, *count=NULL, *send_size=NULL, *recv_size=NULL;
int *partial_send=NULL, *start_pos=NULL, req_len=0, flag=0;
int *recd_from_proc=NULL;
MPI_Aint buftype_extent=0;
size_t byte_size = 0;
OMPI_MPI_OFFSET_TYPE st_loc=-1, end_loc=-1, off=0, done=0, for_next_iter=0;
OMPI_MPI_OFFSET_TYPE size=0, req_off=0, real_size=0, real_off=0, len=0;
OMPI_MPI_OFFSET_TYPE for_curr_iter=0;
char *read_buf=NULL, *tmp_buf=NULL;
MPI_Datatype byte = MPI_BYTE;
opal_datatype_type_size(&byte->super,
&byte_size);
for (i = 0; i < fh->f_size; i++){
if (others_req[i].count) {
st_loc = others_req[i].offsets[0];
end_loc = others_req[i].offsets[0];
break;
}
}
for (i=0;i<fh->f_size;i++){
for(j=0;j< others_req[i].count; j++){
st_loc =
OMPIO_MIN(st_loc, others_req[i].offsets[j]);
end_loc =
OMPIO_MAX(end_loc, (others_req[i].offsets[j] +
others_req[i].lens[j] - 1));
}
}
ntimes = (int)((end_loc - st_loc + mca_fcoll_two_phase_cycle_buffer_size)/
mca_fcoll_two_phase_cycle_buffer_size);
if ((st_loc == -1) && (end_loc == -1)){
ntimes = 0;
}
fh->f_comm->c_coll.coll_allreduce (&ntimes,
&max_ntimes,
1,
MPI_INT,
MPI_MAX,
fh->f_comm,
fh->f_comm->c_coll.coll_allreduce_module);
if (ntimes){
read_buf = (char *) calloc (mca_fcoll_two_phase_cycle_buffer_size,
sizeof(char));
if ( NULL == read_buf ){
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
}
curr_offlen_ptr = (int *)calloc (fh->f_size,
sizeof(int));
if (NULL == curr_offlen_ptr){
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
count = (int *)calloc (fh->f_size,
sizeof(int));
if (NULL == count){
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
partial_send = (int *)calloc(fh->f_size, sizeof(int));
if ( NULL == partial_send ){
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
send_size = (int *)malloc(fh->f_size * sizeof(int));
if (NULL == send_size){
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
recv_size = (int *)malloc(fh->f_size * sizeof(int));
if (NULL == recv_size){
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
recd_from_proc = (int *)calloc(fh->f_size,sizeof(int));
if (NULL == recd_from_proc){
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
start_pos = (int *) calloc(fh->f_size, sizeof(int));
if ( NULL == start_pos ){
ret = OMPI_ERR_OUT_OF_RESOURCE;
return ret;
}
done = 0;
off = st_loc;
for_curr_iter = for_next_iter = 0;
ompi_datatype_type_extent(datatype, &buftype_extent);
for (m=0; m<ntimes; m++) {
size = OMPIO_MIN((unsigned)mca_fcoll_two_phase_cycle_buffer_size, end_loc-st_loc+1-done);
real_off = off - for_curr_iter;
real_size = size + for_curr_iter;
for (i=0; i<fh->f_size; i++) count[i] = send_size[i] = 0;
for_next_iter = 0;
for (i=0; i<fh->f_size; i++) {
if (others_req[i].count) {
start_pos[i] = curr_offlen_ptr[i];
for (j=curr_offlen_ptr[i]; j<others_req[i].count;
j++) {
if (partial_send[i]) {
/* this request may have been partially
satisfied in the previous iteration. */
req_off = others_req[i].offsets[j] +
partial_send[i];
req_len = others_req[i].lens[j] -
partial_send[i];
partial_send[i] = 0;
/* modify the off-len pair to reflect this change */
others_req[i].offsets[j] = req_off;
others_req[i].lens[j] = req_len;
}
else {
req_off = others_req[i].offsets[j];
req_len = others_req[i].lens[j];
}
if (req_off < real_off + real_size) {
count[i]++;
MPI_Address(read_buf+req_off-real_off,
&(others_req[i].mem_ptrs[j]));
send_size[i] += (int)(OMPIO_MIN(real_off + real_size - req_off,
(OMPI_MPI_OFFSET_TYPE)req_len));
if (real_off+real_size-req_off < (OMPI_MPI_OFFSET_TYPE)req_len) {
partial_send[i] = (int) (real_off + real_size - req_off);
if ((j+1 < others_req[i].count) &&
(others_req[i].offsets[j+1] <
real_off+real_size)) {
/* this is the case illustrated in the
figure above. */
for_next_iter = OMPIO_MAX(for_next_iter,
real_off + real_size - others_req[i].offsets[j+1]);
/* max because it must cover requests
from different processes */
}
break;
}
}
else break;
}
curr_offlen_ptr[i] = j;
}
}
flag = 0;
for (i=0; i<fh->f_size; i++)
if (count[i]) flag = 1;
if (flag) {
#if TIME_BREAKDOWN
start_read_time = MPI_Wtime();
#endif
len = size * byte_size;
fh->f_io_array = (mca_io_ompio_io_array_t *)calloc
(1,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_io_array[0].length = len;
fh->f_io_array[0].memory_address =
read_buf+for_curr_iter;
fh->f_num_of_io_entries = 1;
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;
}
}
#if 0
int ii;
printf("%d: len/4 : %lld\n",
fh->f_rank,
len/4);
for (ii = 0; ii < len/4 ;ii++){
printf("%d: read_buf[%d]: %ld\n",
fh->f_rank,
ii,
(int *)read_buf[ii]);
}
#endif
fh->f_num_of_io_entries = 0;
if (NULL != fh->f_io_array) {
free (fh->f_io_array);
fh->f_io_array = NULL;
}
#if TIME_BREAKDOWN
end_read_time = MPI_Wtime();
read_time += (end_read_time - start_read_time);
#endif
}
for_curr_iter = for_next_iter;
for (i=0; i< fh->f_size; i++){
recv_size[i] = 0;
}
two_phase_exchange_data(fh, buf, offset_len,
send_size, start_pos, recv_size, count,
partial_send, recd_from_proc,
contig_access_count,
min_st_offset, fd_size, fd_start, fd_end,
flat_buf, others_req, m, buf_idx,
buftype_extent, striping_unit, aggregator_list);
if (for_next_iter){
tmp_buf = (char *) calloc (for_next_iter, sizeof(char));
memcpy(tmp_buf,
read_buf+real_size-for_next_iter,
for_next_iter);
free(read_buf);
read_buf = (char *)malloc(for_next_iter+mca_fcoll_two_phase_cycle_buffer_size);
memcpy(read_buf, tmp_buf, for_next_iter);
free(tmp_buf);
}
off += size;
done += size;
}
for (i=0; i<fh->f_size; i++) count[i] = send_size[i] = 0;
for (m=ntimes; m<max_ntimes; m++)
two_phase_exchange_data(fh, buf, offset_len, send_size,
start_pos, recv_size, count,
partial_send, recd_from_proc,
contig_access_count,
min_st_offset, fd_size, fd_start, fd_end,
flat_buf, others_req, m, buf_idx,
buftype_extent, striping_unit, aggregator_list);
if (ntimes){
free(read_buf);
read_buf = NULL;
}
if (NULL != curr_offlen_ptr){
free(curr_offlen_ptr);
curr_offlen_ptr = NULL;
}
if (NULL != count){
free(count);
count = NULL;
}
if (NULL != partial_send){
free(partial_send);
partial_send = NULL;
}
if (NULL != send_size){
free(send_size);
send_size = NULL;
}
if (NULL != recv_size){
free(recv_size);
recv_size = NULL;
}
if (NULL != recd_from_proc){
free(recd_from_proc);
recd_from_proc = NULL;
}
if (NULL != start_pos){
free(start_pos);
start_pos = NULL;
}
exit:
return ret;
}
static int two_phase_exch_and_write(mca_io_ompio_file_t *fh,
void *buf,
MPI_Datatype datatype,
mca_io_ompio_access_array_t *others_req,
struct iovec *offset_len,
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,
size_t *buf_idx, int striping_unit,
int *aggregator_list)
{
int i, j, ntimes, max_ntimes, m;
int *curr_offlen_ptr=NULL, *count=NULL, *send_size=NULL, *recv_size=NULL;
int *partial_recv=NULL, *start_pos=NULL, req_len, flag;
int *sent_to_proc=NULL, ret = OMPI_SUCCESS;
int *send_buf_idx=NULL, *curr_to_proc=NULL, *done_to_proc=NULL;
OMPI_MPI_OFFSET_TYPE st_loc=-1, end_loc=-1, off, done;
OMPI_MPI_OFFSET_TYPE size=0, req_off, len;
MPI_Aint buftype_extent;
int hole;
size_t byte_size;
MPI_Datatype byte = MPI_BYTE;
#if DEBUG_ON
int ii,jj;
#endif
char *write_buf=NULL;
opal_datatype_type_size(&byte->super,
&byte_size);
for (i = 0; i < fh->f_size; i++){
if (others_req[i].count) {
st_loc = others_req[i].offsets[0];
end_loc = others_req[i].offsets[0];
break;
}
}
for (i=0;i<fh->f_size;i++){
for(j=0;j< others_req[i].count; j++){
st_loc = OMPIO_MIN(st_loc, others_req[i].offsets[j]);
end_loc = OMPIO_MAX(end_loc, (others_req[i].offsets[j] + others_req[i].lens[j] - 1));
}
}
ntimes = (int) ((end_loc - st_loc + mca_fcoll_two_phase_cycle_buffer_size)/mca_fcoll_two_phase_cycle_buffer_size);
if ((st_loc == -1) && (end_loc == -1)) {
ntimes = 0;
}
fh->f_comm->c_coll.coll_allreduce (&ntimes,
&max_ntimes,
1,
MPI_INT,
MPI_MAX,
fh->f_comm,
fh->f_comm->c_coll.coll_allreduce_module);
if (ntimes){
write_buf = (char *) malloc (mca_fcoll_two_phase_cycle_buffer_size);
if ( NULL == write_buf ){
return OMPI_ERR_OUT_OF_RESOURCE;
}
}
curr_offlen_ptr = (int *) calloc(fh->f_size, sizeof(int));
if ( NULL == curr_offlen_ptr ){
return OMPI_ERR_OUT_OF_RESOURCE;
}
count = (int *) malloc(fh->f_size*sizeof(int));
if ( NULL == count ){
return OMPI_ERR_OUT_OF_RESOURCE;
}
partial_recv = (int *)calloc(fh->f_size, sizeof(int));
if ( NULL == partial_recv ){
return OMPI_ERR_OUT_OF_RESOURCE;
}
send_size = (int *) calloc(fh->f_size,sizeof(int));
if ( NULL == send_size ){
return OMPI_ERR_OUT_OF_RESOURCE;
}
recv_size = (int *) calloc(fh->f_size,sizeof(int));
if ( NULL == recv_size ){
return OMPI_ERR_OUT_OF_RESOURCE;
}
send_buf_idx = (int *) malloc(fh->f_size*sizeof(int));
if ( NULL == send_buf_idx ){
return OMPI_ERR_OUT_OF_RESOURCE;
}
sent_to_proc = (int *) calloc(fh->f_size, sizeof(int));
if ( NULL == sent_to_proc){
return OMPI_ERR_OUT_OF_RESOURCE;
}
curr_to_proc = (int *) malloc(fh->f_size*sizeof(int));
if ( NULL == curr_to_proc ){
return OMPI_ERR_OUT_OF_RESOURCE;
}
done_to_proc = (int *) malloc(fh->f_size*sizeof(int));
if ( NULL == done_to_proc ){
return OMPI_ERR_OUT_OF_RESOURCE;
}
start_pos = (int *) malloc(fh->f_size*sizeof(int));
if ( NULL == start_pos ){
return OMPI_ERR_OUT_OF_RESOURCE;
}
done = 0;
off = st_loc;
ompi_datatype_type_extent(datatype, &buftype_extent);
for (m=0;m <ntimes; m++){
for (i=0; i< fh->f_size; i++) count[i] = recv_size[i] = 0;
size = OMPIO_MIN((unsigned)mca_fcoll_two_phase_cycle_buffer_size,
end_loc-st_loc+1-done);
for (i=0;i<fh->f_size;i++){
if(others_req[i].count){
start_pos[i] = curr_offlen_ptr[i];
for (j=curr_offlen_ptr[i]; j<others_req[i].count; j++) {
if (partial_recv[i]) {
/* this request may have been partially
satisfied in the previous iteration. */
req_off = others_req[i].offsets[j] +
partial_recv[i];
req_len = others_req[i].lens[j] -
partial_recv[i];
partial_recv[i] = 0;
/* modify the off-len pair to reflect this change */
others_req[i].offsets[j] = req_off;
others_req[i].lens[j] = req_len;
}
else {
req_off = others_req[i].offsets[j];
req_len = others_req[i].lens[j];
}
if (req_off < off + size) {
count[i]++;
#if DEBUG_ON
printf("%d: req_off : %lld, off : %lld, size : %lld, count[%d]: %d\n", fh->f_rank,
req_off,
off,
size,i,
count[i]);
#endif
MPI_Address(write_buf+req_off-off,
&(others_req[i].mem_ptrs[j]));
#if DEBUG_ON
printf("%d : mem_ptrs : %ld\n", fh->f_rank,
others_req[i].mem_ptrs[j]);
#endif
recv_size[i] += (int) (OMPIO_MIN(off + size - req_off,
(unsigned)req_len));
if (off+size-req_off < (unsigned)req_len){
partial_recv[i] = (int)(off + size - req_off);
break;
}
}
else break;
}
curr_offlen_ptr[i] = j;
}
}
ret = two_phase_exchage_data(fh, buf, write_buf,
offset_len,send_size,
start_pos,recv_size,off,size,
count, partial_recv, sent_to_proc,
contig_access_count,
min_st_offset,
fd_size, fd_start,
fd_end, flat_buf, others_req,
send_buf_idx, curr_to_proc,
done_to_proc, m, buf_idx,
buftype_extent, striping_unit,
aggregator_list, &hole);
if ( OMPI_SUCCESS != ret ){
goto exit;
}
flag = 0;
for (i=0; i<fh->f_size; i++)
if (count[i]) flag = 1;
if (flag){
#if TIME_BREAKDOWN
start_write_time = MPI_Wtime();
#endif
#if DEBUG_ON
printf("rank : %d enters writing\n", fh->f_rank);
printf("size : %ld, off : %ld\n",size, off);
for (ii=0, jj=0;jj<size;jj+=4, ii++){
printf("%d : write_buf[%d]: %d\n", fh->f_rank, ii,((int *)write_buf[jj]));
}
#endif
len = size * byte_size;
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_io_array[0].length = len;
fh->f_io_array[0].memory_address = write_buf;
fh->f_num_of_io_entries = 1;
#if DEBUG_ON
for (i=0 ; i<fh->f_num_of_io_entries ; i++) {
printf("%d: ADDRESS: %p OFFSET: %ld LENGTH: %d\n",
fh->f_rank,
fh->f_io_array[i].memory_address,
fh->f_io_array[i].offset,
fh->f_io_array[i].length);
}
#endif
if (fh->f_num_of_io_entries){
if (OMPI_SUCCESS != fh->f_fbtl->fbtl_pwritev (fh, NULL)) {
opal_output(1, "WRITE FAILED\n");
return OMPI_ERROR;
}
}
#if TIME_BREAKDOWN
end_write_time = MPI_Wtime();
write_time += (end_write_time - start_write_time);
#endif
}
/***************** DONE WRITING *****************************************/
/****RESET **********************/
fh->f_num_of_io_entries = 0;
if (NULL != fh->f_io_array) {
free (fh->f_io_array);
fh->f_io_array = NULL;
}
off += size;
done += size;
}
for (i=0; i<fh->f_size; i++) count[i] = recv_size[i] = 0;
for (m=ntimes; m<max_ntimes; m++) {
ret = two_phase_exchage_data(fh, buf, write_buf,
offset_len,send_size,
start_pos,recv_size,off,size,
count, partial_recv, sent_to_proc,
contig_access_count,
min_st_offset,
fd_size, fd_start,
fd_end, flat_buf,others_req,
send_buf_idx, curr_to_proc,
done_to_proc, m, buf_idx,
buftype_extent, striping_unit,
aggregator_list, &hole);
if ( OMPI_SUCCESS != ret ){
goto exit;
}
}
exit:
if (ntimes){
if ( NULL != write_buf ){
free(write_buf);
}
}
if ( NULL != curr_offlen_ptr ){
free(curr_offlen_ptr);
}
if ( NULL != count ){
free(count);
}
if ( NULL != partial_recv ){
free(partial_recv);
}
if ( NULL != send_size ){
free(send_size);
}
if ( NULL != recv_size ){
free(recv_size);
}
if ( NULL != sent_to_proc ){
free(sent_to_proc);
}
if ( NULL != start_pos ){
free(start_pos);
}
if ( NULL != send_buf_idx ){
free(send_buf_idx);
}
if ( NULL != curr_to_proc ){
free(curr_to_proc);
}
if ( NULL != done_to_proc ){
free(done_to_proc);
}
return ret;
}
int
mca_fcoll_two_phase_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, i = 0, j = 0, interleave_count = 0, striping_unit = 0;
MPI_Aint recv_buf_addr = 0;
uint32_t iov_count = 0, ti = 0;
struct iovec *decoded_iov = NULL, *temp_iov = NULL, *iov = NULL;
size_t max_data = 0;
long long_max_data = 0, long_total_bytes = 0;
int domain_size=0, *count_my_req_per_proc=NULL, count_my_req_procs = 0;
int count_other_req_procs;
size_t *buf_indices=NULL;
int *aggregator_list = NULL, local_count = 0, local_size = 0;
int two_phase_num_io_procs=1;
OMPI_MPI_OFFSET_TYPE start_offset = 0, end_offset = 0, fd_size = 0;
OMPI_MPI_OFFSET_TYPE *start_offsets=NULL, *end_offsets=NULL;
OMPI_MPI_OFFSET_TYPE *fd_start=NULL, *fd_end=NULL, min_st_offset = 0;
Flatlist_node *flat_buf=NULL;
mca_io_ompio_access_array_t *my_req=NULL, *others_req=NULL;
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
mca_common_ompio_print_entry nentry;
#endif
// if (opal_datatype_is_predefined(&datatype->super)) {
// fh->f_flags = fh->f_flags | OMPIO_CONTIGUOUS_MEMORY;
// }
if (! (fh->f_flags & OMPIO_CONTIGUOUS_MEMORY)) {
ret = fh->f_decode_datatype ((struct mca_io_ompio_file_t *)fh,
datatype,
count,
buf,
&max_data,
&temp_iov,
&iov_count);
if (OMPI_SUCCESS != ret ){
goto exit;
}
recv_buf_addr = (size_t)(buf);
decoded_iov = (struct iovec *) calloc
(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 - recv_buf_addr);
decoded_iov[ti].iov_len = temp_iov[ti].iov_len;
#if DEBUG
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;
}
fh->f_get_num_aggregators (&two_phase_num_io_procs);
if (-1 == two_phase_num_io_procs ){
ret = fh->f_set_aggregator_props ((struct mca_io_ompio_file_t *)fh,
two_phase_num_io_procs,
max_data);
if (OMPI_SUCCESS != ret){
goto exit;
}
two_phase_num_io_procs = fh->f_final_num_aggrs;
}
if (two_phase_num_io_procs > fh->f_size){
two_phase_num_io_procs = fh->f_size;
}
aggregator_list = (int *) calloc (two_phase_num_io_procs, sizeof(int));
if (NULL == aggregator_list){
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
for (i=0; i< two_phase_num_io_procs; i++){
aggregator_list[i] = i * fh->f_size / two_phase_num_io_procs;
}
ret = fh->f_generate_current_file_view ((struct mca_io_ompio_file_t *)fh,
max_data,
&iov,
&local_count);
if (OMPI_SUCCESS != ret){
goto exit;
}
long_max_data = (long) max_data;
ret = fh->f_comm->c_coll.coll_allreduce (&long_max_data,
&long_total_bytes,
1,
MPI_LONG,
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*/
/*Flattened datatype for ompio is in decoded_iov it translated into
flatbuf*/
flat_buf = (Flatlist_node *)calloc(1, 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;
flat_buf->indices = NULL;
flat_buf->blocklens = NULL;
if ( 0 < count ) {
local_size = OMPIO_MAX(1,iov_count/count);
}
else {
local_size = 0;
}
if ( 0 < local_size ) {
flat_buf->indices =
(OMPI_MPI_OFFSET_TYPE *)calloc(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 *)calloc(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;
for (j = 0 ; j < local_size ; ++j) {
flat_buf->indices[j] = (OMPI_MPI_OFFSET_TYPE)(intptr_t)decoded_iov[j].iov_base;
flat_buf->blocklens[j] = decoded_iov[j].iov_len;
}
#if DEBUG
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
printf("%d: total_bytes:%ld, local_count: %d\n",
fh->f_rank, long_total_bytes, local_count);
for (i=0 ; i<local_count ; i++) {
printf("%d: fcoll:two_phase:read_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;
if ( 0 < local_count ) {
end_offset = (OMPI_MPI_OFFSET_TYPE)(intptr_t)iov[local_count-1].iov_base +
(OMPI_MPI_OFFSET_TYPE)(intptr_t)iov[local_count-1].iov_len - 1;
}
else {
end_offset = 0;
}
#if DEBUG
printf("%d: 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 *)calloc
(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 *)calloc
(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,
OMPI_OFFSET_DATATYPE,
start_offsets,
1,
OMPI_OFFSET_DATATYPE,
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,
OMPI_OFFSET_DATATYPE,
end_offsets,
1,
OMPI_OFFSET_DATATYPE,
fh->f_comm,
fh->f_comm->c_coll.coll_allgather_module);
if ( OMPI_SUCCESS != ret ){
goto exit;
}
#if DEBUG
for (i=0;i<fh->f_size;i++){
printf("%d: 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
printf("%d: 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,
two_phase_num_io_procs);
if (OMPI_SUCCESS != ret){
goto exit;
}
#if DEBUG
for (i=0;i<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,
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
printf("%d count_other_req_procs : %d\n",
fh->f_rank,
count_other_req_procs);
#endif
#if OMPIO_FCOLL_WANT_TIME_BREAKDOWN
start_rexch = MPI_Wtime();
#endif
ret = two_phase_read_and_exch(fh,
buf,
datatype,
others_req,
iov,
local_count,
min_st_offset,
fd_size,
fd_start,
fd_end,
flat_buf,
buf_indices,
striping_unit,
two_phase_num_io_procs,
aggregator_list);
if (OMPI_SUCCESS != ret){
goto exit;
}
#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 (isread_aggregator(fh->f_rank,
two_phase_num_io_procs,
aggregator_list)){
nentry.aggregator = 1;
}
else{
nentry.aggregator = 0;
}
nentry.nprocs_for_coll = two_phase_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 (flat_buf != NULL){
if (flat_buf->blocklens != NULL){
free (flat_buf->blocklens);
}
if (flat_buf->indices != NULL){
free (flat_buf->indices);
}
free (flat_buf);
}
free (start_offsets);
free (end_offsets);
free (aggregator_list);
free (fd_start);
free (decoded_iov);
free (buf_indices);
free (count_my_req_per_proc);
free (my_req);
free (others_req);
free (fd_end);
return ret;
}
