mirror/hdf5
2
0
Fork 0
mirror of https://github.com/HDFGroup/hdf5.git synced 2025-03-25 17:00:45 +08:00
Code Issues Packages Projects Releases Wiki Activity

[svn-r5385] Purpose:

Browse Source 
New Feature

Description:
    Currently, only process 0 writes metadata to disk, leading to a potential
    performance bottleneck as the other processors wait for it to catch up.

Solution:
    Rotate the metadata responsibilities among all processes, speading out the
    work.

Platforms tested:
    IRIX64 6.5 (modi4)
This commit is contained in:
Quincey Koziol 2002-05-10 10:37:48 -05:00
parent c43feb3092
commit fcdc05f307
1 changed files with 21 additions and 4 deletions

25
src/H5FDmpio.c

@ -54,6 +54,7 @@ typedef struct H5FD_mpio_t {
MPI_Info info; /*file information */
int mpi_rank; /* This process's rank */
int mpi_size; /* Total number of processes */
int mpi_round; /* Current round robin process (for metadata I/O) */
hbool_t allsame; /*same data for all procs? */
haddr_t eof; /*end-of-file marker */
haddr_t eoa; /*end-of-address marker */
@ -828,6 +829,7 @@ H5FD_mpio_open(const char *name, unsigned flags, hid_t fapl_id,
file->info = fa->info;
file->mpi_rank = mpi_rank;
file->mpi_size = mpi_size;
file->mpi_round = 0; /* Start metadata writes with process 0 */
file->btype = MPI_DATATYPE_NULL;
file->ftype = MPI_DATATYPE_NULL;
@ -1433,10 +1435,10 @@ H5FD_mpio_write(H5FD_t *_file, H5FD_mem_t UNUSED type, hid_t dxpl_id, haddr_t ad
file->old_use_types = use_types_this_time;
file->use_types = 0;
/* Only p0 will do the actual write if all procs in comm write same data */
/* Only p<round> will do the actual write if all procs in comm write same data */
allsame = H5FD_mpio_tas_allsame(_file, FALSE);
if (allsame && H5_mpi_1_metawrite_g) {
if (file->mpi_rank != 0) {
if (file->mpi_rank != file->mpi_round) {
#ifdef H5FDmpio_DEBUG
if (H5FD_mpio_Debug[(int)'w']) {
fprintf(stdout,
@ -1501,11 +1503,26 @@ H5FD_mpio_write(H5FD_t *_file, H5FD_mem_t UNUSED type, hid_t dxpl_id, haddr_t ad
file->eof = HADDR_UNDEF;
done:
/* if only p0 writes, need to boardcast the ret_value to other processes */
/* if only p<round> writes, need to broadcast the ret_value to other processes */
if (allsame && H5_mpi_1_metawrite_g) {
if (MPI_SUCCESS !=
MPI_Bcast(&ret_value, sizeof(ret_value), MPI_BYTE, 0, file->comm))
MPI_Bcast(&ret_value, sizeof(ret_value), MPI_BYTE, file->mpi_round, file->comm))
HRETURN_ERROR(H5E_INTERNAL, H5E_MPI, FAIL, "MPI_Bcast failed");
/* Round-robin rotate to the next process */
file->mpi_round = (++file->mpi_round)%file->mpi_size;
#ifdef QAK
{
int max,min;
MPI_Allreduce(&file->mpi_round, &max, 1, MPI_INT, MPI_MAX, file->comm);
MPI_Allreduce(&file->mpi_round, &min, 1, MPI_INT, MPI_MIN, file->comm);
if(max!=file->mpi_round)
printf("%s: rank=%d, round=%d, max=%d\n",FUNC,file->mpi_rank,file->mpi_round,max);
if(min!=file->mpi_round)
printf("%s: rank=%d, round=%d, min=%d\n",FUNC,file->mpi_rank,file->mpi_round,min);
}
#endif /* QAK */
}
#ifdef H5FDmpio_DEBUG
if (H5FD_mpio_Debug[(int)'t'])