void ADIOI_GEN_WriteStridedColl(ADIO_File fd, const void *buf, int count,
MPI_Datatype datatype, int file_ptr_type,
ADIO_Offset offset, ADIO_Status * status, int
*error_code)
{
/* Uses a generalized version of the extended two-phase method described
in "An Extended Two-Phase Method for Accessing Sections of
Out-of-Core Arrays", Rajeev Thakur and Alok Choudhary,
Scientific Programming, (5)4:301--317, Winter 1996.
http://www.mcs.anl.gov/home/thakur/ext2ph.ps */
ADIOI_Access *my_req;
/* array of nprocs access structures, one for each other process in
* whose file domain this process's request lies */
ADIOI_Access *others_req;
/* array of nprocs access structures, one for each other process
* whose request lies in this process's file domain. */
int i, filetype_is_contig, nprocs, nprocs_for_coll, myrank;
int contig_access_count = 0, interleave_count = 0, buftype_is_contig;
int *count_my_req_per_proc, count_my_req_procs, count_others_req_procs;
ADIO_Offset orig_fp, start_offset, end_offset, fd_size, min_st_offset, off;
ADIO_Offset *offset_list = NULL, *st_offsets = NULL, *fd_start = NULL,
*fd_end = NULL, *end_offsets = NULL;
MPI_Aint *buf_idx = NULL;
ADIO_Offset *len_list = NULL;
int old_error, tmp_error;
if (fd->hints->cb_pfr != ADIOI_HINT_DISABLE) {
/* Cast away const'ness as the below function is used for read
* and write */
ADIOI_IOStridedColl(fd, (char *) buf, count, ADIOI_WRITE, datatype,
file_ptr_type, offset, status, error_code);
return;
}
MPI_Comm_size(fd->comm, &nprocs);
MPI_Comm_rank(fd->comm, &myrank);
/* the number of processes that actually perform I/O, nprocs_for_coll,
* is stored in the hints off the ADIO_File structure
*/
nprocs_for_coll = fd->hints->cb_nodes;
orig_fp = fd->fp_ind;
/* only check for interleaving if cb_write isn't disabled */
if (fd->hints->cb_write != ADIOI_HINT_DISABLE) {
/* For this process's request, calculate the list of offsets and
* lengths in the file and determine the start and end offsets. */
/* Note: end_offset points to the last byte-offset that will be accessed.
* e.g., if start_offset=0 and 100 bytes to be read, end_offset=99 */
ADIOI_Calc_my_off_len(fd, count, datatype, file_ptr_type, offset,
&offset_list, &len_list, &start_offset,
&end_offset, &contig_access_count);
/* each process communicates its start and end offsets to other
* processes. The result is an array each of start and end offsets stored
* in order of process rank. */
st_offsets = (ADIO_Offset *) ADIOI_Malloc(nprocs * 2 * sizeof(ADIO_Offset));
end_offsets = st_offsets + nprocs;
MPI_Allgather(&start_offset, 1, ADIO_OFFSET, st_offsets, 1, ADIO_OFFSET, fd->comm);
MPI_Allgather(&end_offset, 1, ADIO_OFFSET, end_offsets, 1, ADIO_OFFSET, fd->comm);
/* are the accesses of different processes interleaved? */
for (i = 1; i < nprocs; i++)
if ((st_offsets[i] < end_offsets[i - 1]) && (st_offsets[i] <= end_offsets[i]))
interleave_count++;
/* This is a rudimentary check for interleaving, but should suffice
* for the moment. */
}
ADIOI_Datatype_iscontig(datatype, &buftype_is_contig);
if (fd->hints->cb_write == ADIOI_HINT_DISABLE ||
(!interleave_count && (fd->hints->cb_write == ADIOI_HINT_AUTO))) {
/* use independent accesses */
if (fd->hints->cb_write != ADIOI_HINT_DISABLE) {
ADIOI_Free(offset_list);
ADIOI_Free(st_offsets);
}
fd->fp_ind = orig_fp;
ADIOI_Datatype_iscontig(fd->filetype, &filetype_is_contig);
if (buftype_is_contig && filetype_is_contig) {
if (file_ptr_type == ADIO_EXPLICIT_OFFSET) {
off = fd->disp + (ADIO_Offset) (fd->etype_size) * offset;
ADIO_WriteContig(fd, buf, count, datatype,
ADIO_EXPLICIT_OFFSET, off, status, error_code);
} else
ADIO_WriteContig(fd, buf, count, datatype, ADIO_INDIVIDUAL, 0, status, error_code);
} else
ADIO_WriteStrided(fd, buf, count, datatype, file_ptr_type, offset, status, error_code);
return;
}
/* Divide the I/O workload among "nprocs_for_coll" processes. This is
done by (logically) dividing the file into file domains (FDs); each
process may directly access only its own file domain. */
ADIOI_Calc_file_domains(st_offsets, end_offsets, nprocs,
nprocs_for_coll, &min_st_offset,
&fd_start, &fd_end,
fd->hints->min_fdomain_size, &fd_size, fd->hints->striping_unit);
/* calculate what portions of the access requests of this process are
located in what file domains */
ADIOI_Calc_my_req(fd, offset_list, len_list, contig_access_count,
min_st_offset, fd_start, fd_end, fd_size,
nprocs, &count_my_req_procs, &count_my_req_per_proc, &my_req, &buf_idx);
/* based on everyone's my_req, calculate what requests of other
processes lie in this process's file domain.
count_others_req_procs = number of processes whose requests lie in
this process's file domain (including this process itself)
count_others_req_per_proc[i] indicates how many separate contiguous
requests of proc. i lie in this process's file domain. */
ADIOI_Calc_others_req(fd, count_my_req_procs,
count_my_req_per_proc, my_req,
nprocs, myrank, &count_others_req_procs, &others_req);
ADIOI_Free(count_my_req_per_proc);
ADIOI_Free(my_req[0].offsets);
ADIOI_Free(my_req);
/* exchange data and write in sizes of no more than coll_bufsize. */
/* Cast away const'ness for the below function */
ADIOI_Exch_and_write(fd, (char *) buf, datatype, nprocs, myrank,
others_req, offset_list,
len_list, contig_access_count, min_st_offset,
fd_size, fd_start, fd_end, buf_idx, error_code);
/* If this collective write is followed by an independent write,
* it's possible to have those subsequent writes on other processes
* race ahead and sneak in before the read-modify-write completes.
* We carry out a collective communication at the end here so no one
* can start independent i/o before collective I/O completes.
*
* need to do some gymnastics with the error codes so that if something
* went wrong, all processes report error, but if a process has a more
* specific error code, we can still have that process report the
* additional information */
old_error = *error_code;
if (*error_code != MPI_SUCCESS)
*error_code = MPI_ERR_IO;
/* optimization: if only one process performing i/o, we can perform
* a less-expensive Bcast */
#ifdef ADIOI_MPE_LOGGING
MPE_Log_event(ADIOI_MPE_postwrite_a, 0, NULL);
#endif
if (fd->hints->cb_nodes == 1)
MPI_Bcast(error_code, 1, MPI_INT, fd->hints->ranklist[0], fd->comm);
else {
tmp_error = *error_code;
MPI_Allreduce(&tmp_error, error_code, 1, MPI_INT, MPI_MAX, fd->comm);
}
#ifdef ADIOI_MPE_LOGGING
MPE_Log_event(ADIOI_MPE_postwrite_b, 0, NULL);
#endif
#ifdef AGGREGATION_PROFILE
MPE_Log_event(5012, 0, NULL);
#endif
if ((old_error != MPI_SUCCESS) && (old_error != MPI_ERR_IO))
*error_code = old_error;
/* free all memory allocated for collective I/O */
ADIOI_Free(others_req[0].offsets);
ADIOI_Free(others_req[0].mem_ptrs);
ADIOI_Free(others_req);
ADIOI_Free(buf_idx);
ADIOI_Free(offset_list);
ADIOI_Free(st_offsets);
ADIOI_Free(fd_start);
#ifdef HAVE_STATUS_SET_BYTES
if (status) {
MPI_Count bufsize, size;
/* Don't set status if it isn't needed */
MPI_Type_size_x(datatype, &size);
bufsize = size * count;
MPIR_Status_set_bytes(status, datatype, bufsize);
}
/* This is a temporary way of filling in status. The right way is to
keep track of how much data was actually written during collective I/O. */
#endif
fd->fp_sys_posn = -1; /* set it to null. */
#ifdef AGGREGATION_PROFILE
MPE_Log_event(5013, 0, NULL);
#endif
}
void ADIOI_GEN_ReadStridedColl(ADIO_File fd, void *buf, int count,
MPI_Datatype datatype, int file_ptr_type,
ADIO_Offset offset, ADIO_Status *status, int
*error_code)
{
/* Uses a generalized version of the extended two-phase method described
in "An Extended Two-Phase Method for Accessing Sections of
Out-of-Core Arrays", Rajeev Thakur and Alok Choudhary,
Scientific Programming, (5)4:301--317, Winter 1996.
http://www.mcs.anl.gov/home/thakur/ext2ph.ps */
ADIOI_Access *my_req;
/* array of nprocs structures, one for each other process in
whose file domain this process's request lies */
ADIOI_Access *others_req;
/* array of nprocs structures, one for each other process
whose request lies in this process's file domain. */
int i, filetype_is_contig, nprocs, nprocs_for_coll, myrank;
int contig_access_count=0, interleave_count = 0, buftype_is_contig;
int *count_my_req_per_proc, count_my_req_procs, count_others_req_procs;
ADIO_Offset start_offset, end_offset, orig_fp, fd_size, min_st_offset, off;
ADIO_Offset *offset_list = NULL, *st_offsets = NULL, *fd_start = NULL,
*fd_end = NULL, *end_offsets = NULL;
ADIO_Offset *len_list = NULL;
int *buf_idx = NULL;
#ifdef HAVE_STATUS_SET_BYTES
MPI_Count bufsize, size;
#endif
if (fd->hints->cb_pfr != ADIOI_HINT_DISABLE) {
ADIOI_IOStridedColl (fd, buf, count, ADIOI_READ, datatype,
file_ptr_type, offset, status, error_code);
return;
}
MPI_Comm_size(fd->comm, &nprocs);
MPI_Comm_rank(fd->comm, &myrank);
/* number of aggregators, cb_nodes, is stored in the hints */
nprocs_for_coll = fd->hints->cb_nodes;
orig_fp = fd->fp_ind;
/* only check for interleaving if cb_read isn't disabled */
if (fd->hints->cb_read != ADIOI_HINT_DISABLE) {
/* For this process's request, calculate the list of offsets and
lengths in the file and determine the start and end offsets. */
/* Note: end_offset points to the last byte-offset that will be accessed.
e.g., if start_offset=0 and 100 bytes to be read, end_offset=99*/
ADIOI_Calc_my_off_len(fd, count, datatype, file_ptr_type, offset,
&offset_list, &len_list, &start_offset,
&end_offset, &contig_access_count);
#ifdef RDCOLL_DEBUG
for (i=0; i<contig_access_count; i++) {
DBG_FPRINTF(stderr, "rank %d off %lld len %lld\n",
myrank, offset_list[i], len_list[i]);
}
#endif
/* each process communicates its start and end offsets to other
processes. The result is an array each of start and end offsets
stored in order of process rank. */
st_offsets = (ADIO_Offset *) ADIOI_Malloc(nprocs*sizeof(ADIO_Offset));
end_offsets = (ADIO_Offset *) ADIOI_Malloc(nprocs*sizeof(ADIO_Offset));
MPI_Allgather(&start_offset, 1, ADIO_OFFSET, st_offsets, 1,
ADIO_OFFSET, fd->comm);
MPI_Allgather(&end_offset, 1, ADIO_OFFSET, end_offsets, 1,
ADIO_OFFSET, fd->comm);
/* are the accesses of different processes interleaved? */
for (i=1; i<nprocs; i++)
if ((st_offsets[i] < end_offsets[i-1]) &&
(st_offsets[i] <= end_offsets[i]))
interleave_count++;
/* This is a rudimentary check for interleaving, but should suffice
for the moment. */
}
ADIOI_Datatype_iscontig(datatype, &buftype_is_contig);
if (fd->hints->cb_read == ADIOI_HINT_DISABLE
|| (!interleave_count && (fd->hints->cb_read == ADIOI_HINT_AUTO)))
{
/* don't do aggregation */
if (fd->hints->cb_read != ADIOI_HINT_DISABLE) {
ADIOI_Free(offset_list);
ADIOI_Free(len_list);
ADIOI_Free(st_offsets);
ADIOI_Free(end_offsets);
}
fd->fp_ind = orig_fp;
ADIOI_Datatype_iscontig(fd->filetype, &filetype_is_contig);
if (buftype_is_contig && filetype_is_contig) {
if (file_ptr_type == ADIO_EXPLICIT_OFFSET) {
off = fd->disp + (fd->etype_size) * offset;
ADIO_ReadContig(fd, buf, count, datatype, ADIO_EXPLICIT_OFFSET,
off, status, error_code);
}
else ADIO_ReadContig(fd, buf, count, datatype, ADIO_INDIVIDUAL,
0, status, error_code);
}
else ADIO_ReadStrided(fd, buf, count, datatype, file_ptr_type,
offset, status, error_code);
return;
}
/* We're going to perform aggregation of I/O. Here we call
* ADIOI_Calc_file_domains() to determine what processes will handle I/O
* to what regions. We pass nprocs_for_coll into this function; it is
* used to determine how many processes will perform I/O, which is also
* the number of regions into which the range of bytes must be divided.
* These regions are called "file domains", or FDs.
*
* When this function returns, fd_start, fd_end, fd_size, and
* min_st_offset will be filled in. fd_start holds the starting byte
* location for each file domain. fd_end holds the ending byte location.
* min_st_offset holds the minimum byte location that will be accessed.
*
* Both fd_start[] and fd_end[] are indexed by an aggregator number; this
* needs to be mapped to an actual rank in the communicator later.
*
*/
ADIOI_Calc_file_domains(st_offsets, end_offsets, nprocs,
nprocs_for_coll, &min_st_offset,
&fd_start, &fd_end,
fd->hints->min_fdomain_size, &fd_size,
fd->hints->striping_unit);
/* calculate where the portions of the access requests of this process
* are located in terms of the file domains. this could be on the same
* process or on other processes. this function fills in:
* count_my_req_procs - number of processes (including this one) for which
* this process has requests in their file domain
* count_my_req_per_proc - count of requests for each process, indexed
* by rank of the process
* my_req[] - array of data structures describing the requests to be
* performed by each process (including self). indexed by rank.
* buf_idx[] - array of locations into which data can be directly moved;
* this is only valid for contiguous buffer case
*/
ADIOI_Calc_my_req(fd, offset_list, len_list, contig_access_count,
min_st_offset, fd_start, fd_end, fd_size,
nprocs, &count_my_req_procs,
&count_my_req_per_proc, &my_req,
&buf_idx);
/* perform a collective communication in order to distribute the
* data calculated above. fills in the following:
* count_others_req_procs - number of processes (including this
* one) which have requests in this process's file domain.
* count_others_req_per_proc[] - number of separate contiguous
* requests from proc i lie in this process's file domain.
*/
ADIOI_Calc_others_req(fd, count_my_req_procs,
count_my_req_per_proc, my_req,
nprocs, myrank, &count_others_req_procs,
&others_req);
/* my_req[] and count_my_req_per_proc aren't needed at this point, so
* let's free the memory
*/
ADIOI_Free(count_my_req_per_proc);
for (i=0; i<nprocs; i++) {
if (my_req[i].count) {
ADIOI_Free(my_req[i].offsets);
ADIOI_Free(my_req[i].lens);
}
}
ADIOI_Free(my_req);
/* read data in sizes of no more than ADIOI_Coll_bufsize,
* communicate, and fill user buf.
*/
ADIOI_Read_and_exch(fd, buf, datatype, nprocs, myrank,
others_req, offset_list,
len_list, contig_access_count, min_st_offset,
fd_size, fd_start, fd_end, buf_idx, error_code);
if (!buftype_is_contig) ADIOI_Delete_flattened(datatype);
/* free all memory allocated for collective I/O */
for (i=0; i<nprocs; i++) {
if (others_req[i].count) {
ADIOI_Free(others_req[i].offsets);
ADIOI_Free(others_req[i].lens);
ADIOI_Free(others_req[i].mem_ptrs);
}
}
ADIOI_Free(others_req);
ADIOI_Free(buf_idx);
ADIOI_Free(offset_list);
ADIOI_Free(len_list);
ADIOI_Free(st_offsets);
ADIOI_Free(end_offsets);
ADIOI_Free(fd_start);
ADIOI_Free(fd_end);
#ifdef HAVE_STATUS_SET_BYTES
MPI_Type_size_x(datatype, &size);
bufsize = size * count;
MPIR_Status_set_bytes(status, datatype, bufsize);
/* This is a temporary way of filling in status. The right way is to
keep track of how much data was actually read and placed in buf
during collective I/O. */
#endif
fd->fp_sys_posn = -1; /* set it to null. */
}