Merge remote-tracking branch 'nccl/master' into no-target-id
Этот коммит содержится в:
Wenkai Du
+67
-21
@@ -71,45 +71,66 @@ static const float hwLat [3][NCCL_NUM_ALGORITHMS][NCCL_NUM_PROTOCOLS] =
|
||||
};
|
||||
|
||||
// LL128 max BW (per channel) for the different collectives
|
||||
// ncclCollBroadcast, ncclCollReduce, ncclCollAllGather, ncclCollReduceScatter, ncclCollAllReduce
|
||||
static const double ll128MaxBwPerCh[NCCL_NUM_FUNCTIONS] = { 18.8, 12.0, 18.3, 15.2, 16.7 };
|
||||
// ncclFuncBroadcast, ncclFuncReduce, ncclFuncAllGather, ncclFuncReduceScatter, ncclFuncAllReduce
|
||||
static const double ll128MaxBwPerCh[NCCL_NUM_FUNCTIONS] = { 18.8, 12.0, 18.3, 15.2, 16.9 };
|
||||
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} };
|
||||
|
||||
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_MAX_NTHREADS;
|
||||
int simpleDefaultThreads = (ringGraph->speedIntra*ringGraph->nChannels <= PCI_WIDTH) ? 256 : NCCL_SIMPLE_MAX_NTHREADS;
|
||||
comm->maxThreads[NCCL_ALGO_RING][NCCL_PROTO_SIMPLE] =
|
||||
#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
|
||||
getNthreads("NCCL_NTHREADS", ncclParamNthreads(), 4*WARP_SIZE, NCCL_MAX_NTHREADS, simpleDefaultThreads);
|
||||
comm->maxThreads[NCCL_ALGO_TREE][NCCL_PROTO_SIMPLE] = comm->maxThreads[NCCL_ALGO_COLLNET][NCCL_PROTO_SIMPLE] =
|
||||
getNthreads("NCCL_NTHREADS", ncclParamNthreads(), 4*WARP_SIZE, NCCL_MAX_NTHREADS, NCCL_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(), 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] =
|
||||
getNthreads("NCCL_NTHREADS", ncclParamNthreads(), 2*WARP_SIZE, 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
|
||||
comm->maxThreads[NCCL_ALGO_RING][NCCL_PROTO_LL128] = comm->maxThreads[NCCL_ALGO_TREE][NCCL_PROTO_LL128] = comm->maxThreads[NCCL_ALGO_COLLNET][NCCL_PROTO_LL128] =
|
||||
getNthreads("NCCL_LL128_NTHREADS", ncclParamLl128Nthreads(), NCCL_LL128_MAX_NTHREADS/4, NCCL_LL128_MAX_NTHREADS, NCCL_LL128_MAX_NTHREADS);
|
||||
|
||||
if (comm->nRanks <= 1) return ncclSuccess;
|
||||
int nNodes = comm->nNodes;
|
||||
int nRanks = comm->nRanks;
|
||||
if (nRanks <= 1) return ncclSuccess;
|
||||
|
||||
int compCap80 = minCompCap == 80 && maxCompCap == 80 ? 1 : 0;
|
||||
float ppn = (float)comm->nRanks / comm->nNodes; // if ppn < 2, then we are sending/receiving at the same GPU through the NIC, apply some bw discount
|
||||
int cpuArch, cpuVendor, cpuModel;
|
||||
NCCLCHECK(ncclTopoCpuType(comm->topo, &cpuArch, &cpuVendor, &cpuModel));
|
||||
int index2 = nNodes <= 2 ? nNodes-1 : 2;
|
||||
// LL: for single node, we look at GPU type; for multi-node, we look at CPU type
|
||||
int index1 = nNodes == 1 ? compCap80 : cpuVendor == NCCL_TOPO_CPU_VENDOR_AMD ? 1 : 0;
|
||||
double llMaxBw = llMaxBws[index1][index2];
|
||||
double perChMaxTreeBw = perChMaxTreeBws[compCap80][index2];
|
||||
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 };
|
||||
int intraHw[NCCL_NUM_ALGORITHMS], hw[NCCL_NUM_ALGORITHMS];
|
||||
for (int a=0; a<NCCL_NUM_ALGORITHMS; a++) intraHw[a] = graphs[a]->typeIntra == LINK_NVL ? NCCL_HW_NVLINK : NCCL_HW_PCI;
|
||||
for (int a=0; a<NCCL_NUM_ALGORITHMS; a++) hw[a] = comm->nNodes == 1 ? intraHw[a] : NCCL_HW_NET;
|
||||
for (int a=0; a<NCCL_NUM_ALGORITHMS; a++) hw[a] = nNodes == 1 ? intraHw[a] : NCCL_HW_NET;
|
||||
|
||||
for (int coll=0; coll<NCCL_NUM_FUNCTIONS; coll++) {
|
||||
int nsteps = coll == ncclCollAllReduce ? 2*(comm->nRanks-1) :
|
||||
coll == ncclCollReduceScatter || coll == ncclCollAllGather ? comm->nRanks-1 :
|
||||
comm->nRanks;
|
||||
int nInterSteps = coll == ncclCollAllReduce ? 2*(comm->nNodes-1) :
|
||||
coll == ncclCollReduceScatter || coll == ncclCollAllGather ? comm->nNodes-1 :
|
||||
comm->nNodes;
|
||||
int nsteps = coll == ncclFuncAllReduce ? 2*(nRanks-1) :
|
||||
coll == ncclFuncReduceScatter || coll == ncclFuncAllGather ? nRanks-1 :
|
||||
nRanks;
|
||||
int nInterSteps = coll == ncclFuncAllReduce ? 2*(nNodes-1) :
|
||||
coll == ncclFuncReduceScatter || coll == ncclFuncAllGather ? nNodes-1 :
|
||||
nNodes;
|
||||
|
||||
for (int a=0; a<NCCL_NUM_ALGORITHMS; a++) {
|
||||
if (coll != ncclCollAllReduce && a != NCCL_ALGO_RING) continue;
|
||||
if (coll != ncclFuncAllReduce && a != NCCL_ALGO_RING) continue;
|
||||
|
||||
for (int p=0; p<NCCL_NUM_PROTOCOLS; p++) {
|
||||
float speed = comm->nNodes <= 2 || a == NCCL_ALGO_COLLNET ? graphs[a]->speedIntra : graphs[a]->speedInter;
|
||||
float speed = nNodes <= 2 || a == NCCL_ALGO_COLLNET ? graphs[a]->speedIntra : graphs[a]->speedInter;
|
||||
float busBw = graphs[a]->nChannels * speed;
|
||||
|
||||
// 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);
|
||||
double maxTreeBw = comm->nNodes > 2 ?
|
||||
@@ -118,21 +139,29 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
|
||||
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);
|
||||
#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_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_COLLNET && p == NCCL_PROTO_LL) busBw *= 1.0/6.0; // 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
|
||||
|
||||
// Convert bus BW to algorithm BW
|
||||
float ratio = (a != NCCL_ALGO_RING) ? .5 : (1.0 * comm->nRanks) / nsteps;
|
||||
float ratio = (a != NCCL_ALGO_RING) ? .5 : (1.0 * nRanks) / nsteps;
|
||||
comm->bandwidths[coll][a][p] = busBw * ratio;
|
||||
|
||||
comm->latencies[coll][a][p] = baseLat[a][p];
|
||||
float intraLat = hwLat[intraHw[a]][a][p];
|
||||
float interLat = hwLat[NCCL_HW_NET][a][p];
|
||||
if (comm->nNodes > 1 && p == NCCL_PROTO_LL) intraLat *= 1.8;
|
||||
if (nNodes > 1 && p == NCCL_PROTO_LL) intraLat *= 1.8;
|
||||
if (a == NCCL_ALGO_RING) {
|
||||
float lat = hwLat[hw[a]][a][p];
|
||||
if ((coll == ncclCollReduce || coll == ncclCollBroadcast)) {
|
||||
if ((coll == ncclFuncReduce || coll == ncclFuncBroadcast)) {
|
||||
if (ringGraph->sameChannels) {
|
||||
comm->latencies[coll][a][p] += lat;
|
||||
} else {
|
||||
@@ -144,10 +173,10 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
|
||||
}
|
||||
} else if (a == NCCL_ALGO_TREE) {
|
||||
comm->latencies[coll][a][p] +=
|
||||
2 * ((comm->nRanks/comm->nNodes-1) * intraLat + log2i(comm->nNodes) * interLat);
|
||||
2 * ((nRanks/nNodes-1) * intraLat + log2i(nNodes) * interLat);
|
||||
} else {
|
||||
comm->latencies[coll][a][p] +=
|
||||
2 * (comm->nRanks/comm->nNodes-1) * intraLat + interLat;
|
||||
2 * (nRanks/nNodes-1) * intraLat + interLat;
|
||||
}
|
||||
}
|
||||
}
|
||||
@@ -168,6 +197,15 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
|
||||
INFO(NCCL_ENV, "NCCL_ALGO set by environment to %s", algoStr);
|
||||
NCCLCHECK(parseList(algoStr, ncclAlgoStr, NCCL_NUM_ALGORITHMS, algoEnable));
|
||||
}
|
||||
// Disable CollNet if it is not supported
|
||||
if (comm->collNetSupport == 0) {
|
||||
algoEnable[NCCL_ALGO_COLLNET] = 0;
|
||||
// If user has hard set NCCL_ALGO=COLLNET, ignore it
|
||||
if (algoEnable[NCCL_ALGO_RING] == 0 && algoEnable[NCCL_ALGO_TREE] == 0) {
|
||||
algoEnable[NCCL_ALGO_RING] = algoEnable[NCCL_ALGO_TREE] = 1;
|
||||
if (comm->rank == 0) WARN("CollNet is not supported or fails to initialize, ignoring NCCL_ALGO=COLLNET");
|
||||
}
|
||||
}
|
||||
|
||||
for (int c=0; c<NCCL_NUM_FUNCTIONS; c++) for (int a=0; a<NCCL_NUM_ALGORITHMS; a++) for (int p=0; p<NCCL_NUM_PROTOCOLS; p++) {
|
||||
int pEnable = protoEnable[p];
|
||||
@@ -178,7 +216,7 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
|
||||
}
|
||||
if (pEnable == 0) comm->bandwidths[c][a][p] = 0;
|
||||
// Only disable algo for Allreduce since others only have one
|
||||
if (c == ncclCollAllReduce && algoEnable[a] == 0) comm->bandwidths[c][a][p] = 0;
|
||||
if (c == ncclFuncAllReduce && algoEnable[a] == 0) comm->bandwidths[c][a][p] = 0;
|
||||
}
|
||||
|
||||
if (comm->rank == 0) {
|
||||
@@ -214,7 +252,7 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
|
||||
comm->threadThresholds[a][NCCL_PROTO_LL128] = NCCL_LL128_THREAD_THRESHOLD;
|
||||
comm->threadThresholds[a][NCCL_PROTO_SIMPLE] = NCCL_SIMPLE_THREAD_THRESHOLD;
|
||||
}
|
||||
comm->threadThresholds[NCCL_ALGO_RING][NCCL_PROTO_LL] *= comm->nRanks;
|
||||
comm->threadThresholds[NCCL_ALGO_RING][NCCL_PROTO_LL] *= nRanks;
|
||||
|
||||
// Override defaults with user env
|
||||
char* str = getenv("NCCL_THREAD_THRESHOLDS");
|
||||
@@ -263,8 +301,16 @@ ncclResult_t ncclTopoGetAlgoTime(struct ncclInfo* info, int algorithm, int proto
|
||||
*time = -1.0; return ncclSuccess;
|
||||
}
|
||||
int logSize = log2i(info->nBytes>>6);
|
||||
|
||||
#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
|
||||
if (algorithm == NCCL_ALGO_TREE && logSize < 22) bw *= treeCorrectionFactor[protocol][logSize];
|
||||
else if (algorithm == NCCL_ALGO_RING && logSize < 22) bw *= ringCorrectionFactor[protocol][logSize];
|
||||
#else
|
||||
if (algorithm == NCCL_ALGO_TREE && logSize < 23) bw *= treeCorrectionFactor[protocol][logSize];
|
||||
if (info->nChannels != 0) bw = bw / info->comm->nChannels * info->nChannels;
|
||||
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);
|
||||
return ncclSuccess;
|
||||
}
|
||||
|
||||
Ссылка в новой задаче
Block a user