AI/rocm-systems
2
0
Fork 0
Code Issues Pull-Requests Actions Pakete Projekte Releases Wiki Aktivität

2.10.3-1

Quellcode durchsuchen 
Add support for bfloat16.
Add ncclAvg reduction operation.
Improve performance for aggregated operations.
Improve performance for tree.
Improve network error reporting.
Add NCCL_NET parameter to force a specific network.
Add NCCL_IB_QPS_PER_CONNECTION parameter to split IB traffic onto multiple queue pairs.
Fix topology detection error in WSL2.
Fix proxy memory elements affinity (improve alltoall performance).
Fix graph search on cubemesh topologies.
Fix hang in cubemesh during NVB connections.


[ROCm/rccl commit: 7e51592129]
Dieser Commit ist enthalten in:
Ke Wen
2021-07-08 14:12:04 -07:00
Ursprung 3dd3a1ca66
Commit 58a9f19dc3
52 geänderte Dateien mit 3496 neuen und 2469 gelöschten Zeilen
Patch-Datei herunterladen Vergleichs-Datei herunterladen Alle Dateien ausklappen Alle Dateien einklappen
projects/rccl/src/graph/tuning.cc
+16 -14
Datei anzeigen
@@ -60,20 +60,19 @@ static const float baseLat [NCCL_NUM_ALGORITHMS][NCCL_NUM_PROTOCOLS] = { { 4.4,
#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)*/ { .52, 1.25, 28 }, /* Ring (LL/LL128/Simple)*/ { .47, 1.9, 3.4 }, /* CollNet (LL/LL128/Simple)*/ { .5, 1.2, 4.0 } },
{ /* Tree (LL/LL128/Simple)*/ { .52, 1.25, 28 }, /* Ring (LL/LL128/Simple)*/ { .47, 1.9, 3.4 }, /* CollNet (LL/LL128/Simple)*/ { .5, 1.2, 8.0 } },
/* PCI */
{ /* Tree (LL/LL128/Simple)*/ { 1.0, 1.9, 28 }, /* Ring (LL/LL128/Simple)*/ { 1.0, 2.5, 5.7 }, /* CollNet (LL/LL128/Simple)*/ { 1.0, 1.9, 5.5 } },
{ /* Tree (LL/LL128/Simple)*/ { 1.0, 1.9, 28 }, /* Ring (LL/LL128/Simple)*/ { 1.0, 2.5, 5.7 }, /* CollNet (LL/LL128/Simple)*/ { 1.0, 1.9, 8.0 } },
/* NET */
{ /* Tree (LL/LL128/Simple)*/ { 5.0, 8.5, 28 }, /* Ring (LL/LL128/Simple)*/ { 2.7, 4.0, 9.6 }, /* CollNet (LL/LL128/Simple)*/ { 5.0, 5.0, 10.7 } }
};
// 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;
@@ -100,6 +99,8 @@ 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 };
@@ -125,11 +126,10 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
// Various model refinements
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);
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);
if (a == NCCL_ALGO_COLLNET && p != NCCL_PROTO_SIMPLE) busBw = 0; // Oneshot CollNet only supports Simple
// Convert bus BW to algorithm BW
@@ -157,7 +157,7 @@ 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
}
}
}
@@ -266,11 +266,11 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
// factor is not ideal but works quite well. Powers of two, 64 B to 256MB.
static float treeCorrectionFactor[NCCL_NUM_PROTOCOLS][23] = {
{ 1.0, 1.0, 1.0, 1.0, .9, .8, .7, .7, .7, .7, .6, .5, .4, .4, .5, .6, .7, .8, .9, 1.0, 1.0, 1.0, 1.0 },
{ 1.0, 1.0, 1.0, 1.0, 1.0, .9, .8, .8, .8, .7, .6, .6, .6, .5, .6, .6, .7, .7, .8, .9, .9, .92, .92 },
{ 1.0, 1.0, 1.0, 1.0, 1.0, .9, .8, .8, .8, .7, .6, .6, .6, .6, .6, .6, .8, .9, .9, .9, .9, 1.0, 1.0 },
{ .9, .9, .9, .9, .9, .9, .9, .8, .7, .6, .6, .5, .5, .5, .5, .6, .7, .8, .7, .7, .8, .9, .9 }
};
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) {
@@ -281,6 +281,8 @@ ncclResult_t ncclTopoGetAlgoTime(struct ncclInfo* info, int algorithm, int proto
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
*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;
}