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
2023-06-21 20:54:24 -07:00
parent 0c77c66221 ea38312273
commit abd0615351
83 changed files with 5798 additions and 4090 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graph/paths.cc
+35 -49
View File
@@ -273,7 +273,7 @@ ncclResult_t ncclTopoCheckP2p(struct ncclTopoSystem* system, int64_t id1, int64_
struct ncclTopoNode* intermediateNode = path->list[0]->remNode;
if (intermediateNode->type == GPU) {
intermediateIndex = intermediateNode - system->nodes[GPU].nodes;
if (intermediateRank) *intermediateRank = intermediateNode->gpu.rank[0];
if (intermediateRank) *intermediateRank = intermediateNode->gpu.rank;
}
}
@@ -409,7 +409,7 @@ ncclResult_t ncclTopoCheckGdr(struct ncclTopoSystem* system, int64_t busId, int
if (distance == PATH_PXN) {
// In case of PXN, use the intermediate GPU distance instead
int proxyRank, g;
NCCLCHECK(ncclTopoGetIntermediateRank(system, gpu->gpu.rank[0], netDev, &proxyRank));
NCCLCHECK(ncclTopoGetIntermediateRank(system, gpu->gpu.rank, netDev, &proxyRank));
NCCLCHECK(ncclTopoRankToIndex(system, proxyRank, &g));
struct ncclTopoNode* proxyGpu = system->nodes[GPU].nodes+g;
distance = proxyGpu->paths[NET][n].type;
@@ -489,7 +489,7 @@ ncclResult_t ncclTopoGetIntermediateRank(struct ncclTopoSystem* system, int rank
WARN("Could not find intermediate GPU between GPU rank %d and NIC %d", rank, netDev);
return ncclInternalError;
}
*intermediateRank = node->gpu.rank[0];
*intermediateRank = node->gpu.rank;
} else {
*intermediateRank = rank;
}
@@ -563,6 +563,11 @@ ncclResult_t ncclTopoComputePaths(struct ncclTopoSystem* system, struct ncclComm
NCCLCHECK(ncclTopoSetPaths(system->nodes[NET].nodes+n, system));
}
// Set direct paths to NVSwitches.
for (int n=0; n<system->nodes[NVS].count; n++) {
NCCLCHECK(ncclTopoSetPaths(system->nodes[NVS].nodes+n, system));
}
// Update path for GPUs when we don't want to / can't use GPU Direct P2P
for (int g=0; g<system->nodes[GPU].count; g++) {
for (int p=0; p<system->nodes[GPU].count; p++) {
@@ -578,10 +583,10 @@ ncclResult_t ncclTopoComputePaths(struct ncclTopoSystem* system, struct ncclComm
if (comm == NULL) continue;
// Remove GPUs we can't (or don't want to) communicate with through P2P or SHM
struct ncclPeerInfo* dstInfo = comm->peerInfo+system->nodes[GPU].nodes[g].gpu.rank[0];
struct ncclPeerInfo* dstInfo = comm->peerInfo+system->nodes[GPU].nodes[g].gpu.rank;
for (int p=0; p<system->nodes[GPU].count; p++) {
if (p == g) continue;
struct ncclPeerInfo* srcInfo = comm->peerInfo+system->nodes[GPU].nodes[p].gpu.rank[0];
struct ncclPeerInfo* srcInfo = comm->peerInfo+system->nodes[GPU].nodes[p].gpu.rank;
int p2p;
NCCLCHECK(ncclTransports[TRANSPORT_P2P]->canConnect(&p2p, system, NULL, srcInfo, dstInfo));
if (p2p == 0) {
@@ -589,7 +594,7 @@ ncclResult_t ncclTopoComputePaths(struct ncclTopoSystem* system, struct ncclComm
NCCLCHECK(ncclTransports[TRANSPORT_SHM]->canConnect(&shm, system, NULL, srcInfo, dstInfo));
if (shm == 0) {
// Mark this peer as inaccessible. We'll trim it later.
system->nodes[GPU].nodes[p].paths[GPU][g].count = 0;
system->nodes[GPU].nodes[p].paths[GPU][g].type = PATH_NET;
}
}
}
@@ -603,32 +608,20 @@ ncclResult_t ncclTopoComputePaths(struct ncclTopoSystem* system, struct ncclComm
// Check whether we can access the NIC through another NVLink-connected GPU (PXN)
struct ncclTopoNode* gpu = system->nodes[GPU].nodes+g;
if (ncclPxnDisable(comm) != 1) {
int pxnGpu = -1;
for (int p=0; p<system->nodes[GPU].count; p++) {
if (p == g) continue;
int localGpuIndex;
NCCLCHECK(ncclTopoGetLocalGpu(system, system->nodes[NET].nodes[n].id, &localGpuIndex));
if (localGpuIndex != g && localGpuIndex != -1) {
// PXN = PCI + NVLink.
struct ncclTopoNode* peerNode = system->nodes[GPU].nodes+p;
struct ncclTopoNode* peerNode = system->nodes[GPU].nodes+localGpuIndex;
// Only use PXN for NIC n if remote GPU p ...
if (peerNode->paths[NET][n].type > PATH_PXB || // Is connected to the NIC through PCI
peerNode->paths[GPU][g].type > PATH_NVL || // Is connected to us through NVLink
(peerNode->paths[NET][n].bw <= gpu->paths[NET][n].bw && // Has either higher BW to that NIC
gpu->paths[NET][n].type <= PATH_PXB)) // or avoids going through a CPU
continue;
pxnGpu = p;
int netDev;
NCCLCHECK(ncclTopoGetLocalNet(system, peerNode->gpu.rank[0], &netDev));
// To ensure proper balancing, use preferably a local GPU which advertised that NIC as its preferred one.
if (netDev == netNode->id) break;
}
if (pxnGpu != -1) {
if (peerNode->paths[NET][n].type <= PATH_PXB && // Is connected to the NIC through PCI
peerNode->paths[GPU][g].type <= PATH_NVL && // Is connected to us through NVLink
(peerNode->paths[NET][n].bw > gpu->paths[NET][n].bw || // Has either higher BW to that NIC
gpu->paths[NET][n].type > PATH_PXB)) // or avoids going through a CPU
// We can use that GPU as relay to communicate with that NIC.
// Only enabling it in the GPU->NIC direction for now to favor
// receiving locally and sending remotely (consistent with net.cc)
NCCLCHECK(addInterStep(system, GPU, pxnGpu, GPU, g, NET, n));
NCCLCHECK(addInterStep(system, GPU, localGpuIndex, GPU, g, NET, n));
}
}
// Update path when we dont want to / can't use GPU Direct RDMA.
@@ -659,16 +652,11 @@ ncclResult_t ncclTopoTrimSystem(struct ncclTopoSystem* system, struct ncclComm*
domains[g] = g;
ids[g] = gpu->id;
for (int p=0; p<g; p++) {
if (gpu->paths[GPU][p].count > 0) {
if (gpu->paths[GPU][p].type < PATH_NET) {
domains[g] = std::min(domains[g], domains[p]);
}
}
for (int j=0; j<gpu->gpu.nRanksPerGpu; j++ ) {
if (gpu->gpu.rank[j] == comm->rank) {
myDomain = domains[g];
break;
}
}
if (gpu->gpu.rank == comm->rank) myDomain = domains[g];
}
int ngpus = system->nodes[GPU].count;
@@ -732,7 +720,7 @@ ncclResult_t ncclTopoTrimSystem(struct ncclTopoSystem* system, struct ncclComm*
if (allXgmi) system->type |= RCCL_TOPO_XGMI_ALL;
for (int g = 0; g < system->nodes[GPU].count; g++) {
int net;
NCCLCHECK(ncclTopoGetLocalNet(system, system->nodes[GPU].nodes[g].gpu.rank[0], &net));
NCCLCHECK(ncclTopoGetLocalNet(system, system->nodes[GPU].nodes[g].gpu.rank, 0, &net));
NCCLCHECK(ncclTopoCheckGdr(system, system->nodes[GPU].nodes[g].id, net, 1, &gdr));
if (!gdr) break;
}
@@ -742,16 +730,12 @@ ncclResult_t ncclTopoTrimSystem(struct ncclTopoSystem* system, struct ncclComm*
INFO(NCCL_GRAPH, "GDR is available on all GPUs");
}
}
if (rcclParamEnableIntranet()) {
remove = 0;
system->type |= RCCL_TOPO_FORCE_INTRA;
}
comm->localRanks = 0;
for (int n=0; n<system->nodes[GPU].count; n++ ) {
comm->localRanks += system->nodes[GPU].nodes[n].gpu.nRanksPerGpu;
}
if (comm->localRanks == comm->nRanks && remove) {
comm->localRanks = system->nodes[GPU].count;
if (system->nodes[GPU].count == comm->nRanks && remove) {
for (int n=system->nodes[NET].count-1; n>=0; n--)
NCCLCHECK(ncclTopoRemoveNode(system, NET, n));
}
@@ -808,8 +792,14 @@ static int nextPow2(int v) {
ncclResult_t ncclTopoComputeP2pChannels(struct ncclComm* comm) {
/* here we already honor comm->max/minCTAs for p2pnChannels. */
comm->p2pnChannels = std::min(comm->nChannels, (int)ncclParamMaxP2pNChannels());
comm->p2pnChannels = std::max(comm->p2pnChannels, (int)ncclParamMinP2pNChannels());
if (comm->sharedRes->owner != comm) {
comm->p2pnChannels = std::min(comm->nChannels, (int)ncclParamMaxP2pNChannels());
comm->p2pnChannels = std::min(std::max(comm->p2pnChannels, (int)ncclParamMinP2pNChannels()), comm->sharedRes->tpP2pNChannels);
} else {
comm->p2pnChannels = std::min(comm->nChannels, (int)ncclParamMaxP2pNChannels());
comm->p2pnChannels = std::max(comm->p2pnChannels, (int)ncclParamMinP2pNChannels());
}
int minChannels = comm->p2pnChannels;
// We need to loop through all local GPUs to have a global picture
for (int g=0; g<comm->topo->nodes[GPU].count; g++) {
@@ -857,14 +847,10 @@ ncclResult_t ncclTopoGetNvbGpus(struct ncclTopoSystem* system, int rank, int* nr
int nvbGpus = 0;
for (int g=0; g<ngpus; g++) {
struct ncclTopoNode* gpu = system->nodes[GPU].nodes+g;
int j=0;
for ( ; j<gpu->gpu.nRanksPerGpu; j++ ){
if (gpu->gpu.rank[j] == rank) break;
}
if ( j == gpu->gpu.nRanksPerGpu ) continue;
if (gpu->gpu.rank != rank) continue;
for (int p=0; p<ngpus; p++) {
if (gpu->paths[GPU][p].type == PATH_NVB) {
(*ranks)[nvbGpus++] = system->nodes[GPU].nodes[p].gpu.rank[j];
(*ranks)[nvbGpus++] = system->nodes[GPU].nodes[p].gpu.rank;
}
}
}