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

Merge remote-tracking branch 'nccl/master' into 2.10.3

Browse Source
This commit is contained in:
Wenkai Du
2021-07-28 13:27:06 -07:00
parent 229ca88ee6 7e51592129
commit bf2339f93e
53 changed files with 3806 additions and 3023 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graph/tuning.cc
+16 -13
View File
@@ -61,7 +61,7 @@ static const float baseLat [NCCL_NUM_ALGORITHMS][NCCL_NUM_PROTOCOLS] = { { 39.0
#define NCCL_HW_PCI 1
#define NCCL_HW_NET 2
// 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] =
static 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)*/ { 1.2, 1.2, 3.8 } },
/* PCI */
@@ -70,11 +70,10 @@ static const float hwLat [3][NCCL_NUM_ALGORITHMS][NCCL_NUM_PROTOCOLS] =
{ /* 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
// ncclFuncBroadcast, ncclFuncReduce, ncclFuncAllGather, ncclFuncReduceScatter, ncclFuncAllReduce
static const double ll128MaxBwPerCh[NCCL_NUM_FUNCTIONS] = { 18.8, 12.0, 18.3, 15.2, 16.9 };
// LL128 max BW per channel
static const double ll128MaxBwPerCh = 20.0;
static const double llMaxBws[2][3] = { /* Volta-N1/Intel-N2/Intel-N4) */ {39.0, 39.0, 20.4}, /* Ampere-N1/AMD-N2/AMD-N4) */ {87.7, 22.5 /*avg of ring & tree*/, 19.0} };
static const double perChMaxTreeBws[2][3] = { /* Volta (N1/N2/N4) */ {26.5, 18.5, 10.0}, /* Ampere (N1/N2/N4) */ {24.0, 22.5, 16.0} };
static const double perChMaxTreeBws[2][3] = { /* Volta (N1/N2/N4) */ {26.5, 18.5, 10.0}, /* Ampere (N1/N2/N4) */ {24.0, 23.6, 17.8} };
ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCompCap, struct ncclTopoGraph* treeGraph, struct ncclTopoGraph* ringGraph, struct ncclTopoGraph* collNetGraph) {
int simpleDefaultThreads = (ringGraph->speedIntra*ringGraph->nChannels <= PCI_WIDTH) ? 256 : NCCL_SIMPLE_MAX_NTHREADS;
@@ -109,6 +108,8 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
int index1 = nNodes == 1 ? compCap80 : cpuVendor == NCCL_TOPO_CPU_VENDOR_AMD ? 1 : 0;
double llMaxBw = llMaxBws[index1][index2];
double perChMaxTreeBw = perChMaxTreeBws[compCap80][index2];
// De-penalize Tree/Simple latency on Power systems to favor Tree than Ring
if (cpuArch == NCCL_TOPO_CPU_ARCH_POWER) hwLat[NCCL_HW_PCI][NCCL_ALGO_TREE][NCCL_PROTO_SIMPLE] = hwLat[NCCL_HW_PCI][NCCL_ALGO_RING][NCCL_PROTO_SIMPLE];
float ppn = (float)nRanks / nNodes; // if ppn < 2, then we are sending/receiving at the same GPU through the NIC, apply some bw discount
struct ncclTopoGraph* graphs[NCCL_NUM_ALGORITHMS] = { treeGraph, ringGraph, collNetGraph };
@@ -134,22 +135,22 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
// Various model refinements
#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
if (a == NCCL_ALGO_RING && p == NCCL_PROTO_LL) busBw *= 1.0/5.0;
if (a == NCCL_ALGO_RING && p == NCCL_PROTO_LL128) busBw = std::min(busBw * (ppn < 2 ? 0.7 : 0.92 /*120.0/128.0*/), ll128MaxBwPerCh[coll]*graphs[a]->nChannels);
if (a == NCCL_ALGO_RING && p == NCCL_PROTO_LL128) busBw = std::min(busBw * (ppn < 2 ? 0.7 : 0.92 /*120.0/128.0*/), ll128MaxBwPerCh*graphs[a]->nChannels);
double maxTreeBw = comm->nNodes > 2 ?
compCap80 && p == NCCL_PROTO_LL128 ? 105.0 : 80.0 :
compCap80 && p == NCCL_PROTO_LL128 ? 130.0 : 110.0;
if (a == NCCL_ALGO_TREE) busBw = std::min(busBw*.27, comm->nNodes > 1 ? 70.0 : 90.0);
if (a == NCCL_ALGO_TREE && p == NCCL_PROTO_LL) busBw *= 1.0/2.3;
if (a == NCCL_ALGO_TREE && p == NCCL_PROTO_LL128) busBw = std::min(busBw * (comm->nNodes == 1 ? 7.0/9.0 : 0.915 /*120.0/128.0*/), ll128MaxBwPerCh[coll]*graphs[a]->nChannels*7.0/9.0);
if (a == NCCL_ALGO_TREE && p == NCCL_PROTO_LL128) busBw = std::min(busBw * (comm->nNodes == 1 ? 7.0/9.0 : 0.915 /*120.0/128.0*/), ll128MaxBwPerCh*graphs[a]->nChannels*7.0/9.0);
#else
if (compCap80) busBw = std::min(busBw, 235.0f);
if (a == NCCL_ALGO_RING && p == NCCL_PROTO_LL) { busBw = std::min(llMaxBw, busBw * ((nNodes > 1 || coll == ncclFuncAllReduce || coll == ncclFuncReduce) ? 1.0/4.0 : 1.0/3.0)); }
if (a == NCCL_ALGO_RING && p == NCCL_PROTO_LL128) busBw = std::min(busBw * (ppn < 2 ? 0.7 : 0.92 /*120.0/128.0*/), ll128MaxBwPerCh[coll]*graphs[a]->nChannels);
if (a == NCCL_ALGO_RING && p == NCCL_PROTO_LL128) busBw = std::min(busBw * (ppn < 2 ? 0.7 : 0.92 /*120.0/128.0*/), ll128MaxBwPerCh*graphs[a]->nChannels);
if (a == NCCL_ALGO_TREE) busBw = std::min(busBw*.92, graphs[a]->nChannels*perChMaxTreeBw);
if (a == NCCL_ALGO_TREE && p == NCCL_PROTO_LL) busBw = std::min(busBw*1.0/3.8, llMaxBw);
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_TREE && p == NCCL_PROTO_LL128) busBw = std::min(busBw * (nNodes == 1 ? 7.0/9.0 : 120.0/128.0), ll128MaxBwPerCh*graphs[a]->nChannels);
#endif
if (a == NCCL_ALGO_COLLNET && p != NCCL_PROTO_SIMPLE) busBw = 0; // Oneshot CollNet only supports Simple
// Convert bus BW to algorithm BW
@@ -177,7 +178,7 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
2 * ((nRanks/nNodes-1) * intraLat + log2i(nNodes) * interLat);
} else {
comm->latencies[coll][a][p] +=
2 * (nRanks/nNodes-1) * intraLat + interLat;
2 * (std::min(1, (nRanks/nNodes-1)) * intraLat + (nRanks/nNodes-1) * 0.5) + interLat; // Add 0.5 arity serialization latency
}
}
}
@@ -296,7 +297,7 @@ static float ringCorrectionFactor[NCCL_NUM_PROTOCOLS][22] = {
{ 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, .04, .08, .09, .09, .11, .13, .25, .40, .59, .76, .86, 1.0 , 1.0 , 1.0 , 1.0 , 1.0 }
};
ncclResult_t ncclTopoGetAlgoTime(struct ncclInfo* info, int algorithm, int protocol, float* time) {
ncclResult_t ncclTopoGetAlgoTime(struct ncclInfo* info, int algorithm, int protocol, int numPipeOps, float* time) {
float bw = info->comm->bandwidths[info->coll][algorithm][protocol];
float lat = info->comm->latencies[info->coll][algorithm][protocol];
if (bw == 0) {
@@ -313,6 +314,8 @@ ncclResult_t ncclTopoGetAlgoTime(struct ncclInfo* info, int algorithm, int proto
if (algorithm == NCCL_ALGO_RING && protocol == NCCL_PROTO_SIMPLE && info->comm->nNodes > 1
&& info->coll == ncclFuncAllReduce && info->nBytes >= info->comm->nRanks/16.0*65536) lat *= 1.9; // Plateau effect of ring
#endif
*time = lat + (info->nBytes) / (1000 * bw);
// Tree pipelining saves latency in aggregation cases
int latCount = algorithm == NCCL_ALGO_RING ? numPipeOps : DIVUP(numPipeOps, NCCL_MAX_WORK_ELEMENTS);
*time = lat * latCount + (info->nBytes) / (1000 * bw);
return ncclSuccess;
}