AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'nccl/master' into develop

Browse Source
This commit is contained in:
Wenkai Du
2021-04-30 16:57:36 -07:00
parent 4f8e788a61 a46ea10583
commit a4ea1fed5b
80 changed files with 3136 additions and 1675 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graph/tuning.cc
+11 -8
View File
@@ -1,6 +1,6 @@
/*************************************************************************
* Copyright (c) 2016-2020, NVIDIA CORPORATION. All rights reserved.
* Modifications Copyright (c) 2019-2020 Advanced Micro Devices, Inc. All rights reserved.
* Copyright (c) 2016-2021, NVIDIA CORPORATION. All rights reserved.
* Modifications Copyright (c) 2019-2021 Advanced Micro Devices, Inc. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
@@ -63,11 +63,11 @@ static const float baseLat [NCCL_NUM_ALGORITHMS][NCCL_NUM_PROTOCOLS] = { { 39.0
// Tree/Simple is the latency a 256kB chunk, which is ~ base lat + 256k/12GB/s (+ 256k/12GB/s for the network).
static const float hwLat [3][NCCL_NUM_ALGORITHMS][NCCL_NUM_PROTOCOLS] =
{ /* NVLINK */
{ /* Tree (LL/LL128/Simple)*/ { 2.5, 2.5, 5.5 }, /* Ring (LL/LL128/Simple)*/ { 2.5, 2.5, 5 }, /* CollNet (LL/LL128/Simple)*/ { 1.1, 1.1, 3.3 } },
{ /* Tree (LL/LL128/Simple)*/ { 2.5, 2.5, 5.5 }, /* Ring (LL/LL128/Simple)*/ { 2.5, 2.5, 5 }, /* CollNet (LL/LL128/Simple)*/ { 1.2, 1.2, 3.8 } },
/* PCI */
{ /* Tree (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 }, /* Ring (LL/LL128/Simple)*/ { 1.3, 1.3, 1.9 }, /* CollNet (LL/LL128/Simple)*/ { 1.1, 1.1, 1.7 } },
{ /* Tree (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 }, /* Ring (LL/LL128/Simple)*/ { 1.3, 1.3, 1.9 }, /* CollNet (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 } },
/* NET */
{ /* Tree (LL/LL128/Simple)*/ { 28.0, 28.0, 66.0 }, /* Ring (LL/LL128/Simple)*/ { 8.5, 8.5, 19.0 }, /* CollNet (LL/LL128/Simple)*/ { 6.5, 6.5, 14.5 } }
{ /* Tree (LL/LL128/Simple)*/ { 28.0, 28.0, 66.0 }, /* Ring (LL/LL128/Simple)*/ { 8.5, 8.5, 19.0 }, /* CollNet (LL/LL128/Simple)*/ { 9.8, 9.8, 19.5 } }
};
// LL128 max BW (per channel) for the different collectives
@@ -87,8 +87,10 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
getNthreads("NCCL_NTHREADS", ncclParamNthreads(), 4*WARP_SIZE, NCCL_MAX_NTHREADS, NCCL_MAX_NTHREADS);
#else
getNthreads("NCCL_NTHREADS", ncclParamNthreads(), 2*WARP_SIZE, NCCL_SIMPLE_MAX_NTHREADS, simpleDefaultThreads);
comm->maxThreads[NCCL_ALGO_TREE][NCCL_PROTO_SIMPLE] = comm->maxThreads[NCCL_ALGO_COLLNET][NCCL_PROTO_SIMPLE] =
comm->maxThreads[NCCL_ALGO_TREE][NCCL_PROTO_SIMPLE] =
getNthreads("NCCL_NTHREADS", ncclParamNthreads(), 2*WARP_SIZE, NCCL_SIMPLE_MAX_NTHREADS, NCCL_SIMPLE_MAX_NTHREADS);
comm->maxThreads[NCCL_ALGO_COLLNET][NCCL_PROTO_SIMPLE] =
getNthreads("NCCL_NTHREADS", ncclParamNthreads(), NCCL_SIMPLE_MAX_NTHREADS, NCCL_SIMPLE_MAX_NTHREADS, NCCL_SIMPLE_MAX_NTHREADS);
comm->maxThreads[NCCL_ALGO_RING][NCCL_PROTO_LL] = comm->maxThreads[NCCL_ALGO_TREE][NCCL_PROTO_LL] = comm->maxThreads[NCCL_ALGO_COLLNET][NCCL_PROTO_LL] =
getNthreads("NCCL_NTHREADS", ncclParamNthreads(), 2*WARP_SIZE, NCCL_LL_MAX_NTHREADS, NCCL_LL_MAX_NTHREADS);
#endif
@@ -148,8 +150,7 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
if (a == NCCL_ALGO_TREE && p == NCCL_PROTO_LL128) busBw = std::min(busBw * (nNodes == 1 ? 7.0/9.0 : 0.915 /*120.0/128.0*/), ll128MaxBwPerCh[coll]*graphs[a]->nChannels);
#endif
if (a == NCCL_ALGO_COLLNET) busBw *= .9;
if (a == NCCL_ALGO_COLLNET && p == NCCL_PROTO_LL) busBw *= 1.0/2.0; // Take into account that GDR read is disabled on both sides
if (a == NCCL_ALGO_COLLNET && p == NCCL_PROTO_LL128) busBw = 0; // CollNet does not support LL128
if (a == NCCL_ALGO_COLLNET && p != NCCL_PROTO_SIMPLE) busBw = 0; // Oneshot CollNet only supports Simple
// Convert bus BW to algorithm BW
float ratio = (a != NCCL_ALGO_RING) ? .5 : (1.0 * nRanks) / nsteps;
@@ -158,6 +159,7 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
comm->latencies[coll][a][p] = baseLat[a][p];
float intraLat = hwLat[intraHw[a]][a][p];
float interLat = hwLat[NCCL_HW_NET][a][p];
if (nNodes > 1 && p == NCCL_PROTO_LL) intraLat *= 1.8;
if (a == NCCL_ALGO_RING) {
float lat = hwLat[hw[a]][a][p];
if ((coll == ncclFuncReduce || coll == ncclFuncBroadcast)) {
@@ -252,6 +254,7 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
comm->threadThresholds[a][NCCL_PROTO_SIMPLE] = NCCL_SIMPLE_THREAD_THRESHOLD;
}
comm->threadThresholds[NCCL_ALGO_RING][NCCL_PROTO_LL] *= nRanks;
comm->threadThresholds[NCCL_ALGO_COLLNET][NCCL_PROTO_SIMPLE] = 512;
// Override defaults with user env
char* str = getenv("NCCL_THREAD_THRESHOLDS");