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

2.15.1-1

Browse Source 
Add support for H100 (sm90).
Make sure NCCL kernel honor user stream priorities.


[ROCm/rccl commit: da8152e57a]
This commit is contained in:
Sylvain Jeaugey
2022-09-27 02:31:13 -07:00
parent 37ccaf1f82
commit b4bac0d15a
22 changed files with 303 additions and 76 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/graph/paths.cc
+13
View File
@@ -399,6 +399,19 @@ ncclResult_t ncclTopoCheckGdr(struct ncclTopoSystem* system, int64_t busId, int
return ncclSuccess;
}
// Set to 0 to disable the flush on Hopper when using GDR
NCCL_PARAM(NetForceFlush, "NET_FORCE_FLUSH", 1);
// Determine whether we need to flush the GDR recv buffers
ncclResult_t ncclTopoNeedFlush(struct ncclTopoSystem* system, int64_t busId, int* flush) {
int g;
NCCLCHECK(ncclTopoIdToIndex(system, GPU, busId, &g));
struct ncclTopoNode* gpu = system->nodes[GPU].nodes+g;
// Flush is required on Ampere and earlier
*flush = gpu->gpu.cudaCompCap < 90 ? 1 : ncclParamNetForceFlush();
return ncclSuccess;
}
NCCL_PARAM(NetDisableIntra, "NET_DISABLE_INTRA", 0);
// Check whether going through the network would be faster than going through P2P/SHM.
projects/rccl/src/graph/search.cc
+1 -1
View File
@@ -727,7 +727,7 @@ ncclResult_t ncclTopoGetXmlFromGraphs(int ngraphs, struct ncclTopoGraph** graphs
}
float speedArrayIntra[] = { 44.0, 30.0, 22.0, 18.0, 15.0, 12.0, 10.0, 9.0, 7.0, 6.0, 5.0, 4.0, 3.0 };
float speedArrayInter[] = { 48.0, 30.0, 24.0, 22.0, 18.0, 15.0, 12.0, 10.0, 9.0, 7.0, 6.0, 5.0, 4.0, 3.0, 2.4, 1.2, 0.24, 0.12 };
float speedArrayInter[] = { 48.0, 30.0, 28.0, 24.0, 22.0, 18.0, 15.0, 12.0, 10.0, 9.0, 7.0, 6.0, 5.0, 4.0, 3.0, 2.4, 1.2, 0.24, 0.12 };
#define NSPEEDSINTRA (sizeof(speedArrayIntra)/sizeof(float))
#define NSPEEDSINTER (sizeof(speedArrayInter)/sizeof(float))
projects/rccl/src/graph/tuning.cc
+26 -12
View File
@@ -72,10 +72,24 @@ static float hwLat [3][NCCL_NUM_ALGORITHMS][NCCL_NUM_PROTOCOLS] =
/* CollNetDirect (Simple)*/ { 0, 0, 10.7 }, /* CollNetChain (Simple)*/ { 0, 0, 10.7 } }
};
/* Array indexes used below */
#define VOLTA_COMPCAP_IDX 0
#define AMPERE_COMPCAP_IDX 1
#define HOPPER_COMPCAP_IDX 2
// 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, 23.6, 17.8} };
static const double llMaxBws[3][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},
/* Hopper-N1/AMD-N2/AMD-N4) */ {87.7, 22.5 /*avg of ring & tree*/, 19.0}
};
static const double perChMaxTreeBws[3][3] = {
/* Volta (N1/N2/N4) */ {26.5, 18.5, 10.0},
/* Ampere (N1/N2/N4) */ {24.0, 23.6, 17.8},
/* Hopper (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->bwIntra*ringGraph->nChannels <= PCI_BW) ? 256 : NCCL_SIMPLE_MAX_NTHREADS;
@@ -94,14 +108,14 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
int nRanks = comm->nRanks;
if (nRanks <= 1) return ncclSuccess;
int compCap80 = minCompCap == 80 && maxCompCap == 80 ? 1 : 0;
int compCapIndex = (minCompCap == 80 && maxCompCap == 80) ? AMPERE_COMPCAP_IDX : ((minCompCap == 90 && maxCompCap == 90) ? HOPPER_COMPCAP_IDX : VOLTA_COMPCAP_IDX);
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;
int index1 = nNodes == 1 ? compCapIndex : cpuVendor == NCCL_TOPO_CPU_VENDOR_AMD ? 1 : 0;
double llMaxBw = llMaxBws[index1][index2];
double perChMaxTreeBw = perChMaxTreeBws[compCap80][index2];
double perChMaxTreeBw = perChMaxTreeBws[compCapIndex][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
@@ -128,7 +142,7 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
float busBw = graphs[a]->nChannels * bw;
// Various model refinements
if (compCap80) busBw = std::min(busBw, 235.0f);
if (compCapIndex == AMPERE_COMPCAP_IDX) 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*graphs[a]->nChannels);
if (a == NCCL_ALGO_TREE) busBw = std::min(busBw*.92, graphs[a]->nChannels*perChMaxTreeBw);
@@ -136,13 +150,13 @@ 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 : 120.0/128.0), ll128MaxBwPerCh*graphs[a]->nChannels);
if (a == NCCL_ALGO_COLLNET_DIRECT && p != NCCL_PROTO_SIMPLE) busBw = 0; // Not used
if (a == NCCL_ALGO_COLLNET_CHAIN && p != NCCL_PROTO_SIMPLE) busBw = 0; // Not used
if (a == NCCL_ALGO_COLLNET_DIRECT && p == NCCL_PROTO_SIMPLE) {
if (a == NCCL_ALGO_COLLNET_DIRECT && p == NCCL_PROTO_SIMPLE) {
// Collnet+Direct requires all GPUs to have a local NIC to work at full speed
float factor = ppn / (1.0*graphs[a]->nChannels); // GPU/NIC ratio
factor -= (factor-1)/2;
factor -= (factor-1)/2;
busBw /= factor;
}
if (a == NCCL_ALGO_COLLNET_CHAIN && p == NCCL_PROTO_SIMPLE) busBw *= .75;
}
if (a == NCCL_ALGO_COLLNET_CHAIN && p == NCCL_PROTO_SIMPLE) busBw *= .75;
// Convert bus BW to algorithm BW
float ratio = (a != NCCL_ALGO_RING) ? .5 : (1.0 * nRanks) / nsteps;
@@ -212,9 +226,9 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
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];
if (pEnable == 2 && p == NCCL_PROTO_LL128) {
// Enable LL128 by default only on Volta/Ampere+NVLink. Other cases are not tested and may cause silent data corruption.
// Enable LL128 by default only on Volta/Ampere/Hopper+NVLink. Other cases are not tested and may cause silent data corruption.
pEnable = (graphs[a]->typeInter <= PATH_PXB) && graphs[a]->typeIntra <= PATH_NVL &&
((minCompCap == 70 && maxCompCap == 70) || (minCompCap == 80 && maxCompCap == 80)) ? 1 : 0;
((minCompCap == 70 && maxCompCap == 70) || (minCompCap == 80 && maxCompCap == 80) || (minCompCap == 90 && maxCompCap == 90)) ? 1 : 0;
}
if (pEnable == 0) comm->bandwidths[c][a][p] = 0;
// Only disable algo for Allreduce since others only have one
projects/rccl/src/graph/xml.cc
+16 -3
View File
@@ -628,7 +628,7 @@ ncclResult_t ncclTopoGetXmlFromGpu(struct ncclXmlNode* pciNode, nvmlDevice_t nvm
NCCLCHECK(xmlGetSub(gpuNode, "nvlink", &nvlNode));
if (nvlNode == NULL) {
// NVML NVLink detection
int maxNvLinks = (sm < 60) ? 0 : (sm < 70) ? 4 : (sm < 80) ? 6 : 12;
int maxNvLinks = (sm < 60) ? 0 : (sm < 70) ? 4 : (sm < 80) ? 6 : (sm < 90) ? 12 : 18;
if (maxNvLinks > 0 && nvmlDev == NULL) {
WARN("No NVML device handle. Skipping nvlink detection.");
@@ -641,8 +641,21 @@ ncclResult_t ncclTopoGetXmlFromGpu(struct ncclXmlNode* pciNode, nvmlDevice_t nvm
if ((ncclNvmlDeviceGetNvLinkCapability(nvmlDev, l, NVML_NVLINK_CAP_P2P_SUPPORTED, &canP2P) != ncclSuccess) || !canP2P) continue;
// Make sure the Nvlink is up. The previous call should have trained the link.
nvmlEnableState_t isActive;
if ((ncclNvmlDeviceGetNvLinkState(nvmlDev, l, &isActive) != ncclSuccess) || (isActive != NVML_FEATURE_ENABLED)) continue;
nvmlEnableState_t isActive = NVML_FEATURE_DISABLED;
#if CUDART_VERSION >= 11080
if (sm >= 90) {
nvmlFieldValue_t fv;
fv.fieldId = NVML_FI_DEV_NVLINK_GET_STATE;
fv.scopeId = l;
// fv.value will contain NV_FEATURE_ENABLED or NV_FEATURE_DISABLED
if ((ncclNvmlDeviceGetFieldValues(nvmlDev, 1, &fv) == ncclSuccess) && (fv.nvmlReturn == NVML_SUCCESS))
isActive = (nvmlEnableState_t) fv.value.uiVal;
} else /* FALLTHRU to GetNvLinkState if before SM90 */
#endif
{
(void) ncclNvmlDeviceGetNvLinkState(nvmlDev, l, &isActive);
}
if (isActive != NVML_FEATURE_ENABLED) continue;
// Try to figure out what's on the other side of the NVLink
nvmlPciInfo_t remoteProc;