ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность
9780a000ba
openmpi/ompi/mca/coll/tuned/coll_tuned_allreduce.c
Jelena Pjesivac-Grbovic 627533fe4a Adding segmented ring algorithm for Allreduce for commutative operations. 
Algorithm allows user to specify the segment size to be used for computation/communication overlap.
The additional memory requirement for the algorithm is 2 x segment size.
It performed well for (really) large message sizes over MX and it passed intel Allreduce_c and Allreduce_loc_c tests.

This commit was SVN r13832.
2007-02-27 20:32:30 +00:00

1019 строки
45 KiB
C
Исходник Ответственный История

/*
* Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2006 The University of Tennessee and The University
* of Tennessee Research Foundation. All rights
* reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "mpi.h"
#include "ompi/constants.h"
#include "ompi/datatype/datatype.h"
#include "ompi/communicator/communicator.h"
#include "ompi/mca/coll/coll.h"
#include "ompi/mca/coll/base/coll_tags.h"
#include "ompi/mca/pml/pml.h"
#include "ompi/op/op.h"
#include "coll_tuned.h"
#include "coll_tuned_topo.h"
#include "coll_tuned_util.h"
/*
* ompi_coll_tuned_allreduce_intra_nonoverlapping
*
* This function just calls a reduce followed by a broadcast
* both called functions are tuned but they complete sequentially,
* i.e. no additional overlapping
* meaning if the number of segments used is greater than the topo depth
* then once the first segment of data is fully 'reduced' it is not broadcast
* while the reduce continues (cost = cost-reduce + cost-bcast + decision x 3)
*
*/
int
ompi_coll_tuned_allreduce_intra_nonoverlapping(void *sbuf, void *rbuf, int count,
struct ompi_datatype_t *dtype,
struct ompi_op_t *op,
struct ompi_communicator_t *comm)
{
int err;
int rank;
rank = ompi_comm_rank(comm);
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:allreduce_intra_nonoverlapping rank %d", rank));
/* Reduce to 0 and broadcast. */
if (MPI_IN_PLACE == sbuf) {
if (0 == ompi_comm_rank(comm)) {
err = comm->c_coll.coll_reduce (MPI_IN_PLACE, rbuf, count, dtype,
op, 0, comm);
} else {
err = comm->c_coll.coll_reduce (rbuf, NULL, count, dtype, op, 0,
comm);
}
} else {
err = comm->c_coll.coll_reduce (sbuf, rbuf, count, dtype, op, 0, comm);
}
if (MPI_SUCCESS != err) {
return err;
}
return comm->c_coll.coll_bcast (rbuf, count, dtype, 0, comm);
}
/*
* ompi_coll_tuned_allreduce_intra_recursivedoubling
*
* Function: Recursive doubling algorithm for allreduce operation
* Accepts: Same as MPI_Allreduce()
* Returns: MPI_SUCCESS or error code
*
* Description: Implements recursive doubling algorithm for allreduce.
* Original (non-segmented) implementation is used in MPICH-2
* for small and intermediate size messages.
* The algorithm preserves order of operations so it can
* be used both by commutative and non-commutative operations.
*
* Example on 7 nodes:
* Initial state
* # 0 1 2 3 4 5 6
* [0] [1] [2] [3] [4] [5] [6]
* Initial adjustment step for non-power of two nodes.
* old rank 1 3 5 6
* new rank 0 1 2 3
* [0+1] [2+3] [4+5] [6]
* Step 1
* old rank 1 3 5 6
* new rank 0 1 2 3
* [0+1+] [0+1+] [4+5+] [4+5+]
* [2+3+] [2+3+] [6 ] [6 ]
* Step 2
* old rank 1 3 5 6
* new rank 0 1 2 3
* [0+1+] [0+1+] [0+1+] [0+1+]
* [2+3+] [2+3+] [2+3+] [2+3+]
* [4+5+] [4+5+] [4+5+] [4+5+]
* [6 ] [6 ] [6 ] [6 ]
* Final adjustment step for non-power of two nodes
* # 0 1 2 3 4 5 6
* [0+1+] [0+1+] [0+1+] [0+1+] [0+1+] [0+1+] [0+1+]
* [2+3+] [2+3+] [2+3+] [2+3+] [2+3+] [2+3+] [2+3+]
* [4+5+] [4+5+] [4+5+] [4+5+] [4+5+] [4+5+] [4+5+]
* [6 ] [6 ] [6 ] [6 ] [6 ] [6 ] [6 ]
*
*/
int
ompi_coll_tuned_allreduce_intra_recursivedoubling(void *sbuf, void *rbuf,
int count,
struct ompi_datatype_t *dtype,
struct ompi_op_t *op,
struct ompi_communicator_t *comm)
{
int ret, line;
int rank, size, adjsize, remote, distance;
int newrank, newremote, extra_ranks;
char *tmpsend = NULL, *tmprecv = NULL, *tmpswap = NULL, *inplacebuf = NULL;
ptrdiff_t true_lb, true_extent, lb, extent;
ompi_request_t *reqs[2] = {NULL, NULL};
size = ompi_comm_size(comm);
rank = ompi_comm_rank(comm);
OPAL_OUTPUT((ompi_coll_tuned_stream,
"coll:tuned:allreduce_intra_recursivedoubling rank %d", rank));
/* Special case for size == 1 */
if (1 == size) {
if (MPI_IN_PLACE != sbuf) {
ret = ompi_ddt_copy_content_same_ddt(dtype, count, (char*)rbuf, (char*)sbuf);
if (ret < 0) { line = __LINE__; goto error_hndl; }
}
return MPI_SUCCESS;
}
/* Allocate and initialize temporary send buffer */
ret = ompi_ddt_get_extent(dtype, &lb, &extent);
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
ret = ompi_ddt_get_true_extent(dtype, &true_lb, &true_extent);
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
inplacebuf = (char*) malloc(true_extent + (count - 1) * extent);
if (NULL == inplacebuf) { ret = -1; line = __LINE__; goto error_hndl; }
if (MPI_IN_PLACE == sbuf) {
ret = ompi_ddt_copy_content_same_ddt(dtype, count, inplacebuf, (char*)rbuf);
if (ret < 0) { line = __LINE__; goto error_hndl; }
} else {
ret = ompi_ddt_copy_content_same_ddt(dtype, count, inplacebuf, (char*)sbuf);
if (ret < 0) { line = __LINE__; goto error_hndl; }
}
tmpsend = (char*) inplacebuf;
tmprecv = (char*) rbuf;
/* Determine nearest power of two less than or equal to size */
for (adjsize = 0x1; adjsize <= size; adjsize <<= 1); adjsize = adjsize >> 1;
/* Handle non-power-of-two case:
- Even ranks less than 2 * extra_ranks send their data to (rank + 1), and
sets new rank to -1.
- Odd ranks less than 2 * extra_ranks receive data from (rank - 1),
apply appropriate operation, and set new rank to rank/2
- Everyone else sets rank to rank - extra_ranks
*/
extra_ranks = size - adjsize;
if (rank < (2 * extra_ranks)) {
if (0 == (rank % 2)) {
ret = MCA_PML_CALL(send(tmpsend, count, dtype, (rank + 1),
MCA_COLL_BASE_TAG_ALLREDUCE,
MCA_PML_BASE_SEND_STANDARD, comm));
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
newrank = -1;
} else {
ret = MCA_PML_CALL(recv(tmprecv, count, dtype, (rank - 1),
MCA_COLL_BASE_TAG_ALLREDUCE, comm,
MPI_STATUS_IGNORE));
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
/* tmpsend = tmprecv (op) tmpsend */
ompi_op_reduce(op, tmprecv, tmpsend, count, dtype);
newrank = rank >> 1;
}
} else {
newrank = rank - extra_ranks;
}
/* Communication/Computation loop
- Exchange message with remote node.
- Perform appropriate operation taking in account order of operations:
result = value (op) result
*/
for (distance = 0x1; distance < adjsize; distance <<=1) {
if (newrank < 0) break;
/* Determine remote node */
newremote = newrank ^ distance;
remote = (newremote < extra_ranks)?
(newremote * 2 + 1):(newremote + extra_ranks);
/* Exchange the data */
ret = MCA_PML_CALL(irecv(tmprecv, count, dtype, remote,
MCA_COLL_BASE_TAG_ALLREDUCE, comm, &reqs[0]));
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
ret = MCA_PML_CALL(isend(tmpsend, count, dtype, remote,
MCA_COLL_BASE_TAG_ALLREDUCE,
MCA_PML_BASE_SEND_STANDARD, comm, &reqs[1]));
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
ret = ompi_request_wait_all(2, reqs, MPI_STATUSES_IGNORE);
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
/* Apply operation */
if (rank < remote) {
/* tmprecv = tmpsend (op) tmprecv */
ompi_op_reduce(op, tmpsend, tmprecv, count, dtype);
tmpswap = tmprecv;
tmprecv = tmpsend;
tmpsend = tmpswap;
} else {
/* tmpsend = tmprecv (op) tmpsend */
ompi_op_reduce(op, tmprecv, tmpsend, count, dtype);
}
}
/* Handle non-power-of-two case:
- Odd ranks less than 2 * extra_ranks send result from tmpsend to
(rank - 1)
- Even ranks less than 2 * extra_ranks receive result from (rank + 1)
*/
if (rank < (2 * extra_ranks)) {
if (0 == (rank % 2)) {
ret = MCA_PML_CALL(recv(rbuf, count, dtype, (rank + 1),
MCA_COLL_BASE_TAG_ALLREDUCE, comm,
MPI_STATUS_IGNORE));
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
tmpsend = (char*)rbuf;
} else {
ret = MCA_PML_CALL(send(tmpsend, count, dtype, (rank - 1),
MCA_COLL_BASE_TAG_ALLREDUCE,
MCA_PML_BASE_SEND_STANDARD, comm));
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
}
}
/* Ensure that the final result is in rbuf */
if (tmpsend != rbuf) {
ret = ompi_ddt_copy_content_same_ddt(dtype, count, (char*)rbuf, tmpsend);
if (ret < 0) { line = __LINE__; goto error_hndl; }
}
if (NULL != inplacebuf) free(inplacebuf);
return MPI_SUCCESS;
error_hndl:
OPAL_OUTPUT((ompi_coll_tuned_stream, "%s:%4d\tRank %d Error occurred %d\n",
__FILE__, line, rank, ret));
if (NULL != inplacebuf) free(inplacebuf);
return ret;
}
/*
* ompi_coll_tuned_allreduce_intra_ring
*
* Function: Ring algorithm for allreduce operation
* Accepts: Same as MPI_Allreduce()
* Returns: MPI_SUCCESS or error code
*
* Description: Implements ring algorithm for allreduce: the message is
* automatically segmented to segment of size M/N.
* Algorithm requires 2*N - 1 steps.
*
* Limitations: The algorithm DOES NOT preserve order of operations so it
* can be used only for commutative operations.
* In addition, algorithm cannot work if the total count is
* less than size.
* Example on 5 nodes:
* Initial state
* # 0 1 2 3 4
* [00] [10] [20] [30] [40]
* [01] [11] [21] [31] [41]
* [02] [12] [22] [32] [42]
* [03] [13] [23] [33] [43]
* [04] [14] [24] [34] [44]
*
* COMPUTATION PHASE
* Step 0: rank r sends block r to rank (r+1) and receives bloc (r-1)
* from rank (r-1) [with wraparound].
* # 0 1 2 3 4
* [00] [00+10] [20] [30] [40]
* [01] [11] [11+21] [31] [41]
* [02] [12] [22] [22+32] [42]
* [03] [13] [23] [33] [33+43]
* [44+04] [14] [24] [34] [44]
*
* Step 1: rank r sends block (r-1) to rank (r+1) and receives bloc
* (r-2) from rank (r-1) [with wraparound].
* # 0 1 2 3 4
* [00] [00+10] [01+10+20] [30] [40]
* [01] [11] [11+21] [11+21+31] [41]
* [02] [12] [22] [22+32] [22+32+42]
* [33+43+03] [13] [23] [33] [33+43]
* [44+04] [44+04+14] [24] [34] [44]
*
* Step 2: rank r sends block (r-2) to rank (r+1) and receives bloc
* (r-2) from rank (r-1) [with wraparound].
* # 0 1 2 3 4
* [00] [00+10] [01+10+20] [01+10+20+30] [40]
* [01] [11] [11+21] [11+21+31] [11+21+31+41]
* [22+32+42+02] [12] [22] [22+32] [22+32+42]
* [33+43+03] [33+43+03+13] [23] [33] [33+43]
* [44+04] [44+04+14] [44+04+14+24] [34] [44]
*
* Step 3: rank r sends block (r-3) to rank (r+1) and receives bloc
* (r-3) from rank (r-1) [with wraparound].
* # 0 1 2 3 4
* [00] [00+10] [01+10+20] [01+10+20+30] [FULL]
* [FULL] [11] [11+21] [11+21+31] [11+21+31+41]
* [22+32+42+02] [FULL] [22] [22+32] [22+32+42]
* [33+43+03] [33+43+03+13] [FULL] [33] [33+43]
* [44+04] [44+04+14] [44+04+14+24] [FULL] [44]
*
* DISTRIBUTION PHASE: ring ALLGATHER with ranks shifted by 1.
*
*/
int
ompi_coll_tuned_allreduce_intra_ring(void *sbuf, void *rbuf, int count,
struct ompi_datatype_t *dtype,
struct ompi_op_t *op,
struct ompi_communicator_t *comm)
{
int ret, line;
int rank, size, k, recv_from, send_to;
int early_segcount, late_segcount, split_rank, max_segcount;
int block_count, inbi;
size_t typelng;
char *tmpsend = NULL, *tmprecv = NULL;
char *inbuf[2] = {NULL, NULL};
ptrdiff_t true_lb, true_extent, lb, extent;
ptrdiff_t block_offset, max_real_segsize;
ompi_request_t *reqs[2] = {NULL, NULL};
size = ompi_comm_size(comm);
rank = ompi_comm_rank(comm);
OPAL_OUTPUT((ompi_coll_tuned_stream,
"coll:tuned:allreduce_intra_ring rank %d, count %d", rank, count));
/* Special case for size == 1 */
if (1 == size) {
if (MPI_IN_PLACE != sbuf) {
ret = ompi_ddt_copy_content_same_ddt(dtype, count, (char*)rbuf, (char*)sbuf);
if (ret < 0) { line = __LINE__; goto error_hndl; }
}
return MPI_SUCCESS;
}
/* Special case for count less than size - use recursive doubling */
if (count < size) {
OPAL_OUTPUT((ompi_coll_tuned_stream, "coll:tuned:allreduce_ring rank %d/%d, count %d, switching to recursive doubling", rank, size, count));
return (ompi_coll_tuned_allreduce_intra_recursivedoubling(sbuf, rbuf,
count,
dtype, op,
comm));
}
/* Allocate and initialize temporary buffers */
ret = ompi_ddt_get_extent(dtype, &lb, &extent);
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
ret = ompi_ddt_get_true_extent(dtype, &true_lb, &true_extent);
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
ret = ompi_ddt_type_size( dtype, &typelng);
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
/* Determine the number of elements per block and corresponding
block sizes.
The blocks are divided into "early" and "late" ones:
blocks 0 .. (split_rank - 1) are "early" and
blocks (split_rank) .. (size - 1) are "late".
Early blocks are at most 1 element larger than the late ones.
*/
COLL_TUNED_COMPUTE_BLOCKCOUNT( count, size, split_rank,
early_segcount, late_segcount )
max_segcount = early_segcount;
max_real_segsize = true_extent + (max_segcount - 1) * extent;
inbuf[0] = (char*)malloc(max_real_segsize);
if (NULL == inbuf[0]) { ret = -1; line = __LINE__; goto error_hndl; }
if (size > 2) {
inbuf[1] = (char*)malloc(max_real_segsize);
if (NULL == inbuf[1]) { ret = -1; line = __LINE__; goto error_hndl; }
}
/* Handle MPI_IN_PLACE */
if (MPI_IN_PLACE != sbuf) {
ret = ompi_ddt_copy_content_same_ddt(dtype, count, (char*)rbuf, (char*)sbuf);
if (ret < 0) { line = __LINE__; goto error_hndl; }
}
/* Computation loop */
/*
For each of the remote nodes:
- post irecv for block (r-1)
- send block (r)
- in loop for every step k = 2 .. n
- post irecv for block (r + n - k) % n
- wait on block (r + n - k + 1) % n to arrive
- compute on block (r + n - k + 1) % n
- send block (r + n - k + 1) % n
- wait on block (r + 1)
- compute on block (r + 1)
- send block (r + 1) to rank (r + 1)
Note that we must be careful when computing the begining of buffers and
for send operations and computation we must compute the exact block size.
*/
send_to = (rank + 1) % size;
recv_from = (rank + size - 1) % size;
inbi = 0;
/* Initialize first receive from the neighbor on the left */
ret = MCA_PML_CALL(irecv(inbuf[inbi], max_segcount, dtype, recv_from,
MCA_COLL_BASE_TAG_ALLREDUCE, comm, &reqs[inbi]));
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
/* Send first block (my block) to the neighbor on the right */
block_offset = ((rank < split_rank)?
(rank * early_segcount) :
(rank * late_segcount + split_rank));
block_count = ((rank < split_rank)? early_segcount : late_segcount);
tmpsend = ((char*)rbuf) + block_offset * extent;
ret = MCA_PML_CALL(send(tmpsend, block_count, dtype, send_to,
MCA_COLL_BASE_TAG_ALLREDUCE,
MCA_PML_BASE_SEND_STANDARD, comm));
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
for (k = 2; k < size; k++) {
const int prevblock = (rank + size - k + 1) % size;
inbi = inbi ^ 0x1;
/* Post irecv for the current block */
ret = MCA_PML_CALL(irecv(inbuf[inbi], max_segcount, dtype, recv_from,
MCA_COLL_BASE_TAG_ALLREDUCE, comm, &reqs[inbi]));
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
/* Wait on previous block to arrive */
ret = ompi_request_wait(&reqs[inbi ^ 0x1], MPI_STATUS_IGNORE);
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
/* Apply operation on previous block: result goes to rbuf
rbuf[prevblock] = inbuf[inbi ^ 0x1] (op) rbuf[prevblock]
*/
block_offset = ((prevblock < split_rank)?
(prevblock * early_segcount) :
(prevblock * late_segcount + split_rank));
block_count = ((prevblock < split_rank)? early_segcount : late_segcount);
tmprecv = ((char*)rbuf) + block_offset * extent;
ompi_op_reduce(op, inbuf[inbi ^ 0x1], tmprecv, block_count, dtype);
/* send previous block to send_to */
ret = MCA_PML_CALL(send(tmprecv, block_count, dtype, send_to,
MCA_COLL_BASE_TAG_ALLREDUCE,
MCA_PML_BASE_SEND_STANDARD, comm));
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
}
/* Wait on the last block to arrive */
ret = ompi_request_wait(&reqs[inbi], MPI_STATUS_IGNORE);
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
/* Apply operation on the last block (from neighbor (rank + 1)
rbuf[rank+1] = inbuf[inbi] (op) rbuf[rank + 1] */
recv_from = (rank + 1) % size;
block_offset = ((recv_from < split_rank)?
(recv_from * early_segcount) :
(recv_from * late_segcount + split_rank));
block_count = ((recv_from < split_rank)? early_segcount : late_segcount);
tmprecv = ((char*)rbuf) + block_offset * extent;
ompi_op_reduce(op, inbuf[inbi], tmprecv, block_count, dtype);
/* Distribution loop - variation of ring allgather */
send_to = (rank + 1) % size;
recv_from = (rank + size - 1) % size;
for (k = 0; k < size - 1; k++) {
const int recv_data_from = (rank + size - k) % size;
const int send_data_from = (rank + 1 + size - k) % size;
const int send_block_offset =
((send_data_from < split_rank)?
(send_data_from * early_segcount) :
(send_data_from * late_segcount + split_rank));
const int recv_block_offset =
((recv_data_from < split_rank)?
(recv_data_from * early_segcount) :
(recv_data_from * late_segcount + split_rank));
block_count = ((send_data_from < split_rank)?
early_segcount : late_segcount);
tmprecv = (char*)rbuf + recv_block_offset * extent;
tmpsend = (char*)rbuf + send_block_offset * extent;
ret = ompi_coll_tuned_sendrecv(tmpsend, block_count, dtype, send_to,
MCA_COLL_BASE_TAG_ALLREDUCE,
tmprecv, max_segcount, dtype, recv_from,
MCA_COLL_BASE_TAG_ALLREDUCE,
comm, MPI_STATUS_IGNORE, rank);
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl;}
}
if (NULL != inbuf[0]) free(inbuf[0]);
if (NULL != inbuf[1]) free(inbuf[1]);
return MPI_SUCCESS;
error_hndl:
OPAL_OUTPUT((ompi_coll_tuned_stream, "%s:%4d\tRank %d Error occurred %d\n",
__FILE__, line, rank, ret));
if (NULL != inbuf[0]) free(inbuf[0]);
if (NULL != inbuf[1]) free(inbuf[1]);
return ret;
}
/*
* ompi_coll_tuned_allreduce_intra_ring_segmented
*
* Function: Pipelined ring algorithm for allreduce operation
* Accepts: Same as MPI_Allreduce(), segment size
* Returns: MPI_SUCCESS or error code
*
* Description: Implements pipelined ring algorithm for allreduce:
* user supplies suggested segment size for the pipelining of
* reduce operation.
* The segment size determines the number of phases, np, for
* the algorithm execution.
* The message is automatically divided into blocks of
* approximately (count / (np * segcount)) elements.
* At the end of reduction phase, allgather like step is
* executed.
* Algorithm requires (np + 1)*(N - 1) steps.
*
* Limitations: The algorithm DOES NOT preserve order of operations so it
* can be used only for commutative operations.
* In addition, algorithm cannot work if the total size is
* less than size * segment size.
* Example on 3 nodes with 2 phases
* Initial state
* # 0 1 2
* [00a] [10a] [20a]
* [00b] [10b] [20b]
* [01a] [11a] [21a]
* [01b] [11b] [21b]
* [02a] [12a] [22a]
* [02b] [12b] [22b]
*
* COMPUTATION PHASE 0 (a)
* Step 0: rank r sends block ra to rank (r+1) and receives bloc (r-1)a
* from rank (r-1) [with wraparound].
* # 0 1 2
* [00a] [00a+10a] [20a]
* [00b] [10b] [20b]
* [01a] [11a] [11a+21a]
* [01b] [11b] [21b]
* [22a+02a] [12a] [22a]
* [02b] [12b] [22b]
*
* Step 1: rank r sends block (r-1)a to rank (r+1) and receives bloc
* (r-2)a from rank (r-1) [with wraparound].
* # 0 1 2
* [00a] [00a+10a] [00a+10a+20a]
* [00b] [10b] [20b]
* [11a+21a+01a] [11a] [11a+21a]
* [01b] [11b] [21b]
* [22a+02a] [22a+02a+12a] [22a]
* [02b] [12b] [22b]
*
* COMPUTATION PHASE 1 (b)
* Step 0: rank r sends block rb to rank (r+1) and receives bloc (r-1)b
* from rank (r-1) [with wraparound].
* # 0 1 2
* [00a] [00a+10a] [20a]
* [00b] [00b+10b] [20b]
* [01a] [11a] [11a+21a]
* [01b] [11b] [11b+21b]
* [22a+02a] [12a] [22a]
* [22b+02b] [12b] [22b]
*
* Step 1: rank r sends block (r-1)b to rank (r+1) and receives bloc
* (r-2)b from rank (r-1) [with wraparound].
* # 0 1 2
* [00a] [00a+10a] [00a+10a+20a]
* [00b] [10b] [0bb+10b+20b]
* [11a+21a+01a] [11a] [11a+21a]
* [11b+21b+01b] [11b] [21b]
* [22a+02a] [22a+02a+12a] [22a]
* [02b] [22b+01b+12b] [22b]
*
*
* DISTRIBUTION PHASE: ring ALLGATHER with ranks shifted by 1 (same as
* in regular ring algorithm.
*
*/
int
ompi_coll_tuned_allreduce_intra_ring_segmented(void *sbuf, void *rbuf, int count,
struct ompi_datatype_t *dtype,
struct ompi_op_t *op,
struct ompi_communicator_t *comm,
uint32_t segsize)
{
int ret, line;
int rank, size, k, recv_from, send_to;
int early_blockcount, late_blockcount, split_rank;
int segcount, max_segcount;
int num_phases, phase;
int block_count, inbi;
size_t typelng;
char *tmpsend = NULL, *tmprecv = NULL;
char *inbuf[2] = {NULL, NULL};
ptrdiff_t true_lb, true_extent, lb, extent;
ptrdiff_t block_offset, max_real_segsize;
ompi_request_t *reqs[2] = {NULL, NULL};
size = ompi_comm_size(comm);
rank = ompi_comm_rank(comm);
OPAL_OUTPUT((ompi_coll_tuned_stream,
"coll:tuned:allreduce_intra_ring_segmented rank %d, count %d", rank, count));
/* Special case for size == 1 */
if (1 == size) {
if (MPI_IN_PLACE != sbuf) {
ret = ompi_ddt_copy_content_same_ddt(dtype, count, (char*)rbuf, (char*)sbuf);
if (ret < 0) { line = __LINE__; goto error_hndl; }
}
return MPI_SUCCESS;
}
/* Determine segment count based on the suggested segment size */
ret = ompi_ddt_get_extent(dtype, &lb, &extent);
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
ret = ompi_ddt_get_true_extent(dtype, &true_lb, &true_extent);
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
ret = ompi_ddt_type_size( dtype, &typelng);
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
segcount = count;
COLL_TUNED_COMPUTED_SEGCOUNT(segsize, typelng, segcount)
/* Special case for count less than size * segcount - use regular ring */
if (count < size * segcount) {
OPAL_OUTPUT((ompi_coll_tuned_stream, "coll:tuned:allreduce_ring_segmented rank %d/%d, count %d, switching to regular ring", rank, size, count));
return (ompi_coll_tuned_allreduce_intra_ring(sbuf, rbuf, count, dtype, op,
comm));
}
/* Determine the number of phases of the algorithm */
num_phases = count / (size * segcount);
if ((count % (size * segcount) >= size) &&
(count % (size * segcount) > ((size * segcount) / 2))) {
num_phases++;
}
/* Determine the number of elements per block and corresponding
block sizes.
The blocks are divided into "early" and "late" ones:
blocks 0 .. (split_rank - 1) are "early" and
blocks (split_rank) .. (size - 1) are "late".
Early blocks are at most 1 element larger than the late ones.
Note, these blocks will be split into num_phases segments,
out of the largest one will have max_segcount elements.
*/
COLL_TUNED_COMPUTE_BLOCKCOUNT( count, size, split_rank,
early_blockcount, late_blockcount )
COLL_TUNED_COMPUTE_BLOCKCOUNT( early_blockcount, num_phases, inbi,
max_segcount, k)
max_real_segsize = true_extent + (max_segcount - 1) * extent;
/* Allocate and initialize temporary buffers */
inbuf[0] = (char*)malloc(max_real_segsize);
if (NULL == inbuf[0]) { ret = -1; line = __LINE__; goto error_hndl; }
if (size > 2) {
inbuf[1] = (char*)malloc(max_real_segsize);
if (NULL == inbuf[1]) { ret = -1; line = __LINE__; goto error_hndl; }
}
/* Handle MPI_IN_PLACE */
if (MPI_IN_PLACE != sbuf) {
ret = ompi_ddt_copy_content_same_ddt(dtype, count, (char*)rbuf, (char*)sbuf);
if (ret < 0) { line = __LINE__; goto error_hndl; }
}
/* Computation loop: for each phase, repeat ring allreduce computation loop */
for (phase = 0; phase < num_phases; phase ++) {
ptrdiff_t phase_offset;
int early_phase_segcount, late_phase_segcount, split_phase, phase_count;
/*
For each of the remote nodes:
- post irecv for block (r-1)
- send block (r)
To do this, first compute block offset and count, and use block offset
to compute phase offset.
- in loop for every step k = 2 .. n
- post irecv for block (r + n - k) % n
- wait on block (r + n - k + 1) % n to arrive
- compute on block (r + n - k + 1) % n
- send block (r + n - k + 1) % n
- wait on block (r + 1)
- compute on block (r + 1)
- send block (r + 1) to rank (r + 1)
Note that we must be careful when computing the begining of buffers and
for send operations and computation we must compute the exact block size.
*/
send_to = (rank + 1) % size;
recv_from = (rank + size - 1) % size;
inbi = 0;
/* Initialize first receive from the neighbor on the left */
ret = MCA_PML_CALL(irecv(inbuf[inbi], max_segcount, dtype, recv_from,
MCA_COLL_BASE_TAG_ALLREDUCE, comm, &reqs[inbi]));
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
/* Send first block (my block) to the neighbor on the right:
- compute my block and phase offset
- send data */
block_offset = ((rank < split_rank)?
(rank * early_blockcount) :
(rank * late_blockcount + split_rank));
block_count = ((rank < split_rank)? early_blockcount : late_blockcount);
COLL_TUNED_COMPUTE_BLOCKCOUNT(block_count, num_phases, split_phase,
early_phase_segcount, late_phase_segcount)
phase_count = ((phase < split_phase)?
(early_phase_segcount) : (late_phase_segcount));
phase_offset = ((phase < split_phase)?
(phase * early_phase_segcount) :
(phase * late_phase_segcount + split_phase));
tmpsend = ((char*)rbuf) + (block_offset + phase_offset) * extent;
ret = MCA_PML_CALL(send(tmpsend, phase_count, dtype, send_to,
MCA_COLL_BASE_TAG_ALLREDUCE,
MCA_PML_BASE_SEND_STANDARD, comm));
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
for (k = 2; k < size; k++) {
const int prevblock = (rank + size - k + 1) % size;
inbi = inbi ^ 0x1;
/* Post irecv for the current block */
ret = MCA_PML_CALL(irecv(inbuf[inbi], max_segcount, dtype, recv_from,
MCA_COLL_BASE_TAG_ALLREDUCE, comm,
&reqs[inbi]));
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
/* Wait on previous block to arrive */
ret = ompi_request_wait(&reqs[inbi ^ 0x1], MPI_STATUS_IGNORE);
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
/* Apply operation on previous block: result goes to rbuf
rbuf[prevblock] = inbuf[inbi ^ 0x1] (op) rbuf[prevblock]
*/
block_offset = ((prevblock < split_rank)?
(prevblock * early_blockcount) :
(prevblock * late_blockcount + split_rank));
block_count = ((prevblock < split_rank)?
early_blockcount : late_blockcount);
COLL_TUNED_COMPUTE_BLOCKCOUNT(block_count, num_phases, split_phase,
early_phase_segcount, late_phase_segcount)
phase_count = ((phase < split_phase)?
(early_phase_segcount) : (late_phase_segcount));
phase_offset = ((phase < split_phase)?
(phase * early_phase_segcount) :
(phase * late_phase_segcount + split_phase));
tmprecv = ((char*)rbuf) + (block_offset + phase_offset) * extent;
ompi_op_reduce(op, inbuf[inbi ^ 0x1], tmprecv, phase_count, dtype);
/* send previous block to send_to */
ret = MCA_PML_CALL(send(tmprecv, phase_count, dtype, send_to,
MCA_COLL_BASE_TAG_ALLREDUCE,
MCA_PML_BASE_SEND_STANDARD, comm));
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
}
/* Wait on the last block to arrive */
ret = ompi_request_wait(&reqs[inbi], MPI_STATUS_IGNORE);
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl; }
/* Apply operation on the last block (from neighbor (rank + 1)
rbuf[rank+1] = inbuf[inbi] (op) rbuf[rank + 1] */
recv_from = (rank + 1) % size;
block_offset = ((recv_from < split_rank)?
(recv_from * early_blockcount) :
(recv_from * late_blockcount + split_rank));
block_count = ((recv_from < split_rank)?
early_blockcount : late_blockcount);
COLL_TUNED_COMPUTE_BLOCKCOUNT(block_count, num_phases, split_phase,
early_phase_segcount, late_phase_segcount)
phase_count = ((phase < split_phase)?
(early_phase_segcount) : (late_phase_segcount));
phase_offset = ((phase < split_phase)?
(phase * early_phase_segcount) :
(phase * late_phase_segcount + split_phase));
tmprecv = ((char*)rbuf) + (block_offset + phase_offset) * extent;
ompi_op_reduce(op, inbuf[inbi], tmprecv, phase_count, dtype);
}
/* Distribution loop - variation of ring allgather */
send_to = (rank + 1) % size;
recv_from = (rank + size - 1) % size;
for (k = 0; k < size - 1; k++) {
const int recv_data_from = (rank + size - k) % size;
const int send_data_from = (rank + 1 + size - k) % size;
const int send_block_offset =
((send_data_from < split_rank)?
(send_data_from * early_blockcount) :
(send_data_from * late_blockcount + split_rank));
const int recv_block_offset =
((recv_data_from < split_rank)?
(recv_data_from * early_blockcount) :
(recv_data_from * late_blockcount + split_rank));
block_count = ((send_data_from < split_rank)?
early_blockcount : late_blockcount);
tmprecv = (char*)rbuf + recv_block_offset * extent;
tmpsend = (char*)rbuf + send_block_offset * extent;
ret = ompi_coll_tuned_sendrecv(tmpsend, block_count, dtype, send_to,
MCA_COLL_BASE_TAG_ALLREDUCE,
tmprecv, early_blockcount, dtype, recv_from,
MCA_COLL_BASE_TAG_ALLREDUCE,
comm, MPI_STATUS_IGNORE, rank);
if (MPI_SUCCESS != ret) { line = __LINE__; goto error_hndl;}
}
if (NULL != inbuf[0]) free(inbuf[0]);
if (NULL != inbuf[1]) free(inbuf[1]);
return MPI_SUCCESS;
error_hndl:
OPAL_OUTPUT((ompi_coll_tuned_stream, "%s:%4d\tRank %d Error occurred %d\n",
__FILE__, line, rank, ret));
if (NULL != inbuf[0]) free(inbuf[0]);
if (NULL != inbuf[1]) free(inbuf[1]);
return ret;
}
/*
* Linear functions are copied from the BASIC coll module
* they do not segment the message and are simple implementations
* but for some small number of nodes and/or small data sizes they
* are just as fast as tuned/tree based segmenting operations
* and as such may be selected by the decision functions
* These are copied into this module due to the way we select modules
* in V1. i.e. in V2 we will handle this differently and so will not
* have to duplicate code.
* GEF Oct05 after asking Jeff.
*/
/* copied function (with appropriate renaming) starts here */
/*
* allreduce_intra
*
* Function: - allreduce using other MPI collectives
* Accepts: - same as MPI_Allreduce()
* Returns: - MPI_SUCCESS or error code
*/
int
ompi_coll_tuned_allreduce_intra_basic_linear(void *sbuf, void *rbuf, int count,
struct ompi_datatype_t *dtype,
struct ompi_op_t *op,
struct ompi_communicator_t *comm)
{
int err;
int rank;
rank = ompi_comm_rank(comm);
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:allreduce_intra_basic_linear rank %d", rank));
/* Reduce to 0 and broadcast. */
if (MPI_IN_PLACE == sbuf) {
if (0 == ompi_comm_rank(comm)) {
err = ompi_coll_tuned_reduce_intra_basic_linear (MPI_IN_PLACE, rbuf, count, dtype, op, 0, comm);
} else {
err = ompi_coll_tuned_reduce_intra_basic_linear(rbuf, NULL, count, dtype, op, 0, comm);
}
} else {
err = ompi_coll_tuned_reduce_intra_basic_linear(sbuf, rbuf, count, dtype, op, 0, comm);
}
if (MPI_SUCCESS != err) {
return err;
}
return ompi_coll_tuned_bcast_intra_basic_linear(rbuf, count, dtype, 0, comm);
}
/* copied function (with appropriate renaming) ends here */
/* The following are used by dynamic and forced rules */
/* publish details of each algorithm and if its forced/fixed/locked in */
/* as you add methods/algorithms you must update this and the query/map routines */
/* this routine is called by the component only */
/* this makes sure that the mca parameters are set to their initial values and perms */
/* module does not call this they call the forced_getvalues routine instead */
int ompi_coll_tuned_allreduce_intra_check_forced_init (coll_tuned_force_algorithm_mca_param_indices_t *mca_param_indices)
{
int rc, max_alg = 5, requested_alg;
ompi_coll_tuned_forced_max_algorithms[ALLREDUCE] = max_alg;
rc = mca_base_param_reg_int (&mca_coll_tuned_component.super.collm_version,
"allreduce_algorithm_count",
"Number of allreduce algorithms available",
false, true, max_alg, NULL);
mca_param_indices->algorithm_param_index
= mca_base_param_reg_int( &mca_coll_tuned_component.super.collm_version,
"allreduce_algorithm",
"Which allreduce algorithm is used. Can be locked down to any of: 0 ignore, 1 basic linear, 2 nonoverlapping (tuned reduce + tuned bcast), 3 recursive doubling, 4 ring, 5 segmented ring",
false, false, 0, NULL);
mca_base_param_lookup_int( mca_param_indices->algorithm_param_index, &(requested_alg));
if( requested_alg > max_alg ) {
if( 0 == ompi_comm_rank( MPI_COMM_WORLD ) ) {
opal_output( 0, "Allreduce algorithm #%d is not available (range [0..%d]). Switching back to ignore(0)\n",
requested_alg, max_alg );
}
mca_base_param_set_int( mca_param_indices->algorithm_param_index, 0);
}
mca_param_indices->segsize_param_index
= mca_base_param_reg_int( &mca_coll_tuned_component.super.collm_version,
"allreduce_algorithm_segmentsize",
"Segment size in bytes used by default for allreduce algorithms. Only has meaning if algorithm is forced and supports segmenting. 0 bytes means no segmentation.",
false, false, 0, NULL);
mca_param_indices->tree_fanout_param_index
= mca_base_param_reg_int( &mca_coll_tuned_component.super.collm_version,
"allreduce_algorithm_tree_fanout",
"Fanout for n-tree used for allreduce algorithms. Only has meaning if algorithm is forced and supports n-tree topo based operation.",
false, false, ompi_coll_tuned_init_tree_fanout, /* get system wide default */
NULL);
mca_param_indices->chain_fanout_param_index
= mca_base_param_reg_int( &mca_coll_tuned_component.super.collm_version,
"allreduce_algorithm_chain_fanout",
"Fanout for chains used for allreduce algorithms. Only has meaning if algorithm is forced and supports chain topo based operation.",
false, false,
ompi_coll_tuned_init_chain_fanout, /* get system wide default */
NULL);
return (MPI_SUCCESS);
}
int ompi_coll_tuned_allreduce_intra_do_forced(void *sbuf, void *rbuf, int count,
struct ompi_datatype_t *dtype,
struct ompi_op_t *op,
struct ompi_communicator_t *comm)
{
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:allreduce_intra_do_forced selected algorithm %d, segment size %d",
comm->c_coll_selected_data->user_forced[ALLREDUCE].algorithm,
comm->c_coll_selected_data->user_forced[ALLREDUCE].segsize));
switch (comm->c_coll_selected_data->user_forced[ALLREDUCE].algorithm) {
case (0): return ompi_coll_tuned_allreduce_intra_dec_fixed (sbuf, rbuf, count, dtype, op, comm);
case (1): return ompi_coll_tuned_allreduce_intra_basic_linear (sbuf, rbuf, count, dtype, op, comm);
case (2): return ompi_coll_tuned_allreduce_intra_nonoverlapping (sbuf, rbuf, count, dtype, op, comm);
case (3): return ompi_coll_tuned_allreduce_intra_recursivedoubling (sbuf, rbuf, count, dtype, op, comm);
case (4): return ompi_coll_tuned_allreduce_intra_ring (sbuf, rbuf, count, dtype, op, comm);
case (5): return ompi_coll_tuned_allreduce_intra_ring_segmented (sbuf, rbuf, count, dtype, op, comm, comm->c_coll_selected_data->user_forced[ALLREDUCE].segsize);
default:
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:allreduce_intra_do_forced attempt to select algorithm %d when only 0-%d is valid?",
comm->c_coll_selected_data->user_forced[ALLREDUCE].algorithm,
ompi_coll_tuned_forced_max_algorithms[ALLREDUCE]));
return (MPI_ERR_ARG);
} /* switch */
}
int ompi_coll_tuned_allreduce_intra_do_this(void *sbuf, void *rbuf, int count,
struct ompi_datatype_t *dtype,
struct ompi_op_t *op,
struct ompi_communicator_t *comm,
int algorithm, int faninout, int segsize)
{
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:allreduce_intra_do_this algorithm %d topo fan in/out %d segsize %d",
algorithm, faninout, segsize));
switch (algorithm) {
case (0): return ompi_coll_tuned_allreduce_intra_dec_fixed (sbuf, rbuf, count, dtype, op, comm);
case (1): return ompi_coll_tuned_allreduce_intra_basic_linear (sbuf, rbuf, count, dtype, op, comm);
case (2): return ompi_coll_tuned_allreduce_intra_nonoverlapping (sbuf, rbuf, count, dtype, op, comm);
case (3): return ompi_coll_tuned_allreduce_intra_recursivedoubling (sbuf, rbuf, count, dtype, op, comm);
case (4): return ompi_coll_tuned_allreduce_intra_ring (sbuf, rbuf, count, dtype, op, comm);
case (5): return ompi_coll_tuned_allreduce_intra_ring_segmented (sbuf, rbuf, count, dtype, op, comm, segsize);
default:
OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:allreduce_intra_do_this attempt to select algorithm %d when only 0-%d is valid?",
algorithm, ompi_coll_tuned_forced_max_algorithms[ALLREDUCE]));
return (MPI_ERR_ARG);
} /* switch */
}
Показать git blame Копировать постоянную ссылку