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
2022-10-20 15:40:03 +00:00
parent bc8ef779df 99c28f2e75
commit 4f0e223db4
51 changed files with 2139 additions and 1081 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graph/connect.cc
+20 -15
View File
@@ -48,13 +48,15 @@ ncclResult_t ncclTopoPreset(struct ncclComm* comm,
struct ncclChannel* channel = comm->channels+c;
channel->ring.prev = channel->ring.next = -1;
channel->tree.up = -1;
channel->collnetChain.up = -1;
for (int i=0; i<NCCL_MAX_TREE_ARITY; i++) channel->tree.down[i] = -1;
channel->collTree.out = -1;
channel->collTree.headRank = -1;
channel->collTree.nHeads = 0;
channel->collTree.shift = 0;
for (int i=0; i<NCCL_MAX_DIRECT_ARITY; i++) channel->collTree.up[i] = -1;
for (int i=0; i<NCCL_MAX_DIRECT_ARITY; i++) channel->collTree.down[i] = -1;
for (int i=0; i<NCCL_MAX_TREE_ARITY; i++) channel->collnetChain.down[i] = -1;
channel->collnetDirect.out = -1;
channel->collnetDirect.headRank = -1;
channel->collnetDirect.nHeads = 0;
channel->collnetDirect.shift = 0;
for (int i=0; i<NCCL_MAX_DIRECT_ARITY; i++) channel->collnetDirect.up[i] = -1;
for (int i=0; i<NCCL_MAX_DIRECT_ARITY; i++) channel->collnetDirect.down[i] = -1;
int* ringIntra = ringGraph->intra+c*localRanks;
int* treeIntra = treeGraph->intra+c*localRanks;
@@ -76,6 +78,8 @@ ncclResult_t ncclTopoPreset(struct ncclComm* comm,
topoRanks->treeToChild1[c] = treeIntra[child1Index];
channel->tree.up = i == 0 ? -1 : treeIntra[i-1];
channel->tree.down[0] = i == localRanks-1 ? -1 : treeIntra[i+1];
channel->collnetChain.up = i == 0 ? comm->nRanks : treeIntra[i-1];
channel->collnetChain.down[0] = i == localRanks-1 ? -1 : treeIntra[i+1];
}
}
topoRanks->ringPrev[c] = channel->ring.prev;
@@ -571,13 +575,13 @@ static ncclResult_t connectCollNet(struct ncclComm* comm, struct ncclTopoGraph*
int nDown = 0;
for (int i=0; i<nHeads; i++) {
if (rank == heads[i]) { // is head
channel->collTree.headRank = i; // Mark the index for deciding offset in the CUDA kernel
channel->collTree.out = comm->nRanks; // Set root of collTree to id nranks
channel->collnetDirect.headRank = i; // Mark the index for deciding offset in the CUDA kernel
channel->collnetDirect.out = comm->nRanks; // Set root of collnetDirect to id nranks
int* collNetIntra = collNetGraph->intra+i*localRanks;
sprintf(line+strlen(line), "down ");
for (int r=0; r<localRanks; r++) {
if (collNetIntra[r] == rank) continue;
channel->collTree.down[nDown++] = collNetIntra[r]; // connect to all peers
channel->collnetDirect.down[nDown++] = collNetIntra[r]; // connect to all peers
sprintf(line+strlen(line), " %d ", collNetIntra[r]);
}
sprintf(line+strlen(line), "nDown %d ", nDown);
@@ -589,15 +593,16 @@ static ncclResult_t connectCollNet(struct ncclComm* comm, struct ncclTopoGraph*
sprintf(line+strlen(line), "up ");
for (int h=0; h<nHeads; h++) {
if (rank == heads[h]) continue;
channel->collTree.up[nUp++] = heads[h];
channel->collnetDirect.up[nUp++] = heads[h];
sprintf(line+strlen(line), " %d ", heads[h]);
}
channel->collTree.nHeads = nHeads;
channel->collTree.shift = (rank%localRanks)%nHeads; // Shift by intraRank so that leaves don't send to same head simultaneously
channel->collTree.depth = (nUp == 0 && nDown == 0) ? 1 : 2;
channel->collnetDirect.nHeads = nHeads;
channel->collnetDirect.shift = (rank%localRanks)%nHeads; // Shift by intraRank so that leaves don't send to same head simultaneously
channel->collnetDirect.depth = (nUp == 0 && nDown == 0) ? 1 : 2;
sprintf(line+strlen(line), "nUp %d nHeads %d ", nUp, nHeads);
sprintf(line+strlen(line), "headRank %d out %d shift %d", channel->collTree.headRank, channel->collTree.out, channel->collTree.shift);
sprintf(line+strlen(line), "headRank %d out %d shift %d", channel->collnetDirect.headRank, channel->collnetDirect.out, channel->collnetDirect.shift);
INFO(NCCL_GRAPH, "%s", line);
channel->collnetChain.depth = comm->nRanks/comm->nNodes;
}
free(heads);
return ncclSuccess;
@@ -685,7 +690,7 @@ ncclResult_t ncclTopoPostset(struct ncclComm* comm, int* firstRanks, int* treePa
// Setup CollNet
if (comm->collNetSupport == 1) {
// Add more channels to saturate intra-node bandwidth, except the 1 PPN case
if (collNetGraph->speedIntra > collNetGraph->speedInter && comm->nRanks > comm->nNodes) {
if (collNetGraph->bwIntra > collNetGraph->bwInter && comm->nRanks > comm->nNodes) {
int collNetNchannels = std::min(MAXCHANNELS, nChannels+nChannels/2);
nChannels = comm->nChannels = copyChannels(comm, nChannels, collNetNchannels, ringPrev, ringNext);
}
src/graph/paths.cc
+77 -26
View File
@@ -45,7 +45,7 @@ static ncclResult_t ncclTopoSetPaths(struct ncclTopoNode* baseNode, struct ncclT
struct ncclTopoLinkList* basePath;
NCCLCHECK(getPath(system, baseNode, baseNode->type, baseNode->id, &basePath));
basePath->count = 0;
basePath->width = LOC_WIDTH;
basePath->bw = LOC_BW;
basePath->type = PATH_LOC;
while (nodeList.count) {
@@ -62,8 +62,13 @@ static ncclResult_t ncclTopoSetPaths(struct ncclTopoNode* baseNode, struct ncclT
}
struct ncclTopoLinkList* remPath;
NCCLCHECK(getPath(system, remNode, baseNode->type, baseNode->id, &remPath));
float width = std::min(path->width, link->width);
if (remPath->width < width) {
float bw = std::min(path->bw, link->bw);
// allow routing through a GPU only as 1 hop
if (node != baseNode && node->type == GPU &&
(ncclParamNvbDisable() || link->type != LINK_NVL || remNode->type != GPU || path->count > 1)) continue;
if ((remPath->bw == 0 || remPath->count > path->count) && remPath->bw < bw) {
// Find reverse link
for (int l=0; l<remNode->nlinks; l++) {
if (remNode->links[l].remNode == node) {
@@ -79,7 +84,7 @@ static ncclResult_t ncclTopoSetPaths(struct ncclTopoNode* baseNode, struct ncclT
// Copy the rest of the path
for (int i=0; i<path->count; i++) remPath->list[i+1] = path->list[i];
remPath->count = path->count + 1;
remPath->width = width;
remPath->bw = bw;
// Start with path type = link type. PATH and LINK types are supposed to match.
// Don't consider LINK_NET as we only care about the NIC->GPU path.
@@ -128,9 +133,9 @@ static void printNodePaths(struct ncclTopoSystem* system, struct ncclTopoNode* n
sprintf(line+offset, "--%s->%s/%lX", topoLinkTypeStr[link->type], topoNodeTypeStr[remNode->type], remNode->id);
offset = strlen(line);
}
INFO(NCCL_GRAPH, "%s (%f)", line, node->paths[t][n].width);
INFO(NCCL_GRAPH, "%s (%f)", line, node->paths[t][n].bw);
#else
sprintf(line+offset, "%s/%lX (%d/%f/%s) ", topoNodeTypeStr[t], system->nodes[t].nodes[n].id, node->paths[t][n].count, node->paths[t][n].width, topoPathTypeStr[node->paths[t][n].type]);
sprintf(line+offset, "%s/%lX (%d/%f/%s) ", topoNodeTypeStr[t], system->nodes[t].nodes[n].id, node->paths[t][n].count, node->paths[t][n].bw, topoPathTypeStr[node->paths[t][n].type]);
offset = strlen(line);
#endif
}
@@ -184,7 +189,7 @@ static ncclResult_t addInterStep(struct ncclTopoSystem* system, int tx, int ix,
srcNode->paths[t2][i2].count = l;
srcNode->paths[t2][i2].type = std::max(srcNode->paths[tx][ix].type, cpuNode->paths[t2][i2].type);
if (tx == GPU) srcNode->paths[t2][i2].type = PATH_PXN;
srcNode->paths[t2][i2].width = std::min(srcNode->paths[tx][ix].width, cpuNode->paths[t2][i2].width);
srcNode->paths[t2][i2].bw = std::min(srcNode->paths[tx][ix].bw, cpuNode->paths[t2][i2].bw);
return ncclSuccess;
}
@@ -419,6 +424,40 @@ ncclResult_t ncclTopoCheckGdr(struct ncclTopoSystem* system, int64_t busId, int
return ncclSuccess;
}
NCCL_PARAM(NetDisableIntra, "NET_DISABLE_INTRA", 1);
// Check whether going through the network would be faster than going through P2P/SHM.
ncclResult_t ncclTopoCheckNet(struct ncclTopoSystem* system, int64_t id1, int64_t id2, int* net) {
if (ncclParamNetDisableIntra() == 1) {
*net = 0;
return ncclSuccess;
}
*net = 1;
// First check the current GPU-to-GPU speed.
int g1, g2;
if (ncclTopoIdToIndex(system, GPU, id1, &g1) != ncclSuccess ||
ncclTopoIdToIndex(system, GPU, id2, &g2) != ncclSuccess) {
return ncclSuccess;
}
struct ncclTopoNode* gpu1 = system->nodes[GPU].nodes+g1;
struct ncclTopoNode* gpu2 = system->nodes[GPU].nodes+g2;
float speed = gpu1->paths[GPU][g2].bw;
// Now check the speed each GPU can access the network through PXB or better
float netSpeed1 = 0, netSpeed2 = 0;
for (int n=0; n<system->nodes[NET].count; n++) {
struct ncclTopoLinkList* path = gpu1->paths[NET]+n;
if (path->type <= PATH_PXB && path->bw > netSpeed1) netSpeed1 = path->bw;
path = gpu2->paths[NET]+n;
if (path->type <= PATH_PXB && path->bw > netSpeed2) netSpeed2 = path->bw;
}
if (netSpeed1 > speed && netSpeed2 > speed) return ncclSuccess;
*net = 0;
return ncclSuccess;
}
ncclResult_t ncclTopoGetIntermediateRank(struct ncclTopoSystem* system, int rank, int netDev, int* intermediateRank) {
// Get GPU and NET
int n, g;
@@ -496,17 +535,23 @@ ncclResult_t ncclTopoComputePaths(struct ncclTopoSystem* system, struct ncclComm
// Remove everything in case we're re-computing
for (int t=0; t<NCCL_TOPO_NODE_TYPES; t++) ncclTopoRemovePathType(system, t);
// Set direct paths from/to CPUs. We need them in many cases.
// Set direct paths to CPUs. We need them in many cases.
for (int c=0; c<system->nodes[CPU].count; c++) {
NCCLCHECK(ncclTopoSetPaths(system->nodes[CPU].nodes+c, system));
}
// Set direct paths from/to GPUs.
// Set direct paths to GPUs.
for (int g=0; g<system->nodes[GPU].count; g++) {
// Compute paths to GPU g
NCCLCHECK(ncclTopoSetPaths(system->nodes[GPU].nodes+g, system));
}
// Update path when we don't want to / can't use GPU Direct P2P
// Set direct paths to NICs.
for (int n=0; n<system->nodes[NET].count; n++) {
NCCLCHECK(ncclTopoSetPaths(system->nodes[NET].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++) {
int p2p;
NCCLCHECK(ncclTopoCheckP2p(system, system->nodes[GPU].nodes[p].id, system->nodes[GPU].nodes[g].id, &p2p, NULL, NULL));
@@ -519,31 +564,32 @@ ncclResult_t ncclTopoComputePaths(struct ncclTopoSystem* system, struct ncclComm
}
if (comm == NULL) continue;
// Remove GPUs we can't talk to because of containers.
// 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];
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];
int shm;
NCCLCHECK(ncclTransports[TRANSPORT_SHM]->canConnect(&shm, system, NULL, srcInfo, dstInfo));
int p2p;
NCCLCHECK(ncclTransports[TRANSPORT_P2P]->canConnect(&p2p, system, NULL, srcInfo, dstInfo));
if (shm == 0 && p2p == 0) {
// Mark this peer as inaccessible. We'll trim it later.
system->nodes[GPU].nodes[p].paths[GPU][g].count = 0;
if (p2p == 0) {
int shm;
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;
}
}
}
}
// Set direct paths from/to NICs.
// Update paths for NICs (no GPU Direct, PXN, ...)
for (int n=0; n<system->nodes[NET].count; n++) {
struct ncclTopoNode* netNode = system->nodes[NET].nodes+n;
NCCLCHECK(ncclTopoSetPaths(netNode, system));
for (int g=0; g<system->nodes[GPU].count; g++) {
// 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 && gpu->paths[NET][n].type > PATH_PXB) {
if (ncclPxnDisable(comm) != 1) {
int pxnGpu = -1;
for (int p=0; p<system->nodes[GPU].count; p++) {
@@ -551,7 +597,12 @@ ncclResult_t ncclTopoComputePaths(struct ncclTopoSystem* system, struct ncclComm
// PXN = PCI + NVLink.
struct ncclTopoNode* peerNode = system->nodes[GPU].nodes+p;
if (peerNode->paths[NET][n].type > PATH_PXB || peerNode->paths[GPU][g].type > PATH_NVL) continue;
// 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;
@@ -630,15 +681,15 @@ ncclResult_t ncclTopoTrimSystem(struct ncclTopoSystem* system, struct ncclComm*
for (int j = 0; j < system->nodes[NET].count; j ++) {
if (i == j) continue;
if (system->nodes[NET].nodes[i].net.asic == system->nodes[NET].nodes[j].net.asic) {
if (system->nodes[NET].nodes[i].net.width > system->nodes[NET].nodes[j].net.width)
system->nodes[NET].nodes[j].net.width = 0;
if (system->nodes[NET].nodes[i].net.bw > system->nodes[NET].nodes[j].net.bw)
system->nodes[NET].nodes[j].net.bw = 0;
}
}
}
do {
int n;
for (n=0; n<system->nodes[NET].count; n++) {
if (system->nodes[NET].nodes[n].net.width == 0) break;
if (system->nodes[NET].nodes[n].net.bw == 0) break;
}
if (n<system->nodes[NET].count) {
NCCLCHECK(ncclTopoRemoveNode(system, NET, n));
@@ -716,8 +767,8 @@ static ncclResult_t ncclTopoGetNchannels(struct ncclTopoSystem* system, int g /*
// Local rank
path = system->nodes[GPU].nodes[peer].paths[GPU]+g;
if (path->type == PATH_NVL) {
float nvlWidth = ncclTopoXGMISpeed(system->nodes[GPU].nodes[g].gpu.cudaCompCap);
*nChannels = 2*std::max(1, (int)(path->width / nvlWidth));
float nvlBw = ncclTopoXGMISpeed(system->nodes[GPU].nodes[g].gpu.gcn);
*nChannels = 2*std::max(1, (int)(path->bw / nvlBw));
} else {
*nChannels = 2;
}
src/graph/rome_models.cc
+2 -2
View File
@@ -721,7 +721,7 @@ newchannel:
} while (str[offset++] != 0);
end:
graph->nChannels = nChannels;
graph->speedIntra = graph->speedInter = system->maxWidth;
graph->bwIntra = graph->bwInter = system->maxBw;
if (graph->id == 1) {
for (int i=0; i<graph->nChannels; i++) {
int net;
@@ -1031,7 +1031,7 @@ static ncclResult_t parseRomeSystem(struct ncclTopoSystem* system, struct rcclRo
}
if (!link->remNode) continue;
if (link->type != LINK_NVL) continue;
romeTopo->connMatrix[i*romeTopo->nGpus+n] = link->width/ncclTopoXGMISpeed(node->gpu.gcn);
romeTopo->connMatrix[i*romeTopo->nGpus+n] = link->bw/ncclTopoXGMISpeed(node->gpu.gcn);
count ++;
}
if (romeTopo->nLinks < count) romeTopo->nLinks = count;
src/graph/search.cc
+88 -88
View File
@@ -13,39 +13,39 @@
#include <sys/time.h>
#include "rome_models.h"
// Initialize system->maxWidth. This is the per-channel (i.e. per-SM)
// max speed.
static float getMaxWidth(struct ncclTopoSystem* system, struct ncclTopoNode* gpu, int type) {
float maxWidth = 0.0;
// Initialize system->maxBw. This is the per-channel (i.e. per-SM)
// max bw.
static float getMaxBw(struct ncclTopoSystem* system, struct ncclTopoNode* gpu, int type) {
float maxBw = 0.0;
for (int i=0; i<system->nodes[type].count; i++) {
struct ncclTopoLinkList* path = gpu->paths[type]+i;
float width = path->width;
float bw = path->bw;
if (path->count == 0) continue;
maxWidth = std::max(maxWidth, width);
maxBw = std::max(maxBw, bw);
}
return maxWidth;
return maxBw;
}
static float getTotalWidth(struct ncclTopoSystem* system, struct ncclTopoNode* gpu) {
float nvlinkWidth = 0.0, pciWidth = 0.0;
static float getTotalBw(struct ncclTopoSystem* system, struct ncclTopoNode* gpu) {
float nvlinkBw = 0.0, pciBw = 0.0;
for (int l=0; l<gpu->nlinks; l++) {
struct ncclTopoLink* link = gpu->links+l;
if (link->type == LINK_NVL) nvlinkWidth += link->width;
if (link->type == LINK_PCI) pciWidth = link->width;
if (link->type == LINK_NVL) nvlinkBw += link->bw;
if (link->type == LINK_PCI) pciBw = link->bw;
}
return std::max(pciWidth, nvlinkWidth);
return std::max(pciBw, nvlinkBw);
}
ncclResult_t ncclTopoSearchInit(struct ncclTopoSystem* system) {
system->maxWidth = 0.0;
system->totalWidth = 0.0;
system->maxBw = 0.0;
system->totalBw = 0.0;
int inter = system->nodes[NET].count;
if (inter == 0 && system->nodes[GPU].count == 1) {
system->maxWidth = LOC_WIDTH;
system->maxBw = LOC_BW;
return ncclSuccess;
}
for (int g=0; g<system->nodes[GPU].count; g++) {
struct ncclTopoNode* gpu = system->nodes[GPU].nodes+g;
system->maxWidth = std::max(system->maxWidth, getMaxWidth(system, gpu, inter ? NET : GPU));
system->totalWidth = std::max(system->totalWidth, getTotalWidth(system, gpu));
system->maxBw = std::max(system->maxBw, getMaxBw(system, gpu, inter ? NET : GPU));
system->totalBw = std::max(system->totalBw, getTotalBw(system, gpu));
}
return ncclSuccess;
}
@@ -65,8 +65,8 @@ static ncclResult_t findRevLink(struct ncclTopoNode* node1, struct ncclTopoNode*
// This is unfortunately needed since manipulating floats often results in rounding errors.
#define SUB_ROUND(a, b) (a = roundf((a-b)*1000)/1000)
static ncclResult_t followPath(struct ncclTopoLinkList* path, struct ncclTopoNode* start, int maxSteps, float speed, int* steps) {
float pciSpeed = speed;
static ncclResult_t followPath(struct ncclTopoLinkList* path, struct ncclTopoNode* start, int maxSteps, float bw, int* steps) {
float pciBw = bw;
for (int step=0; step<path->count; step++) {
struct ncclTopoNode* node = path->list[step]->remNode;
if (node->type == CPU) {
@@ -74,7 +74,7 @@ static ncclResult_t followPath(struct ncclTopoLinkList* path, struct ncclTopoNod
if (path->type == PATH_PHB && start->type == GPU &&
node->cpu.arch == NCCL_TOPO_CPU_ARCH_X86 &&
node->cpu.vendor == NCCL_TOPO_CPU_VENDOR_INTEL) {
pciSpeed = INTEL_P2P_OVERHEAD(speed);
pciBw = INTEL_P2P_OVERHEAD(bw);
}
}
}
@@ -83,19 +83,19 @@ static ncclResult_t followPath(struct ncclTopoLinkList* path, struct ncclTopoNod
for (int step=0; step<maxSteps; step++) {
struct ncclTopoLink* link = path->list[step];
struct ncclTopoLink* revLink = NULL;
float fwSpeed = link->type == LINK_PCI ? pciSpeed : speed;
float revSpeed = 0;
float fwBw = link->type == LINK_PCI ? pciBw : bw;
float revBw = 0;
if (link->remNode->type == GPU && link->remNode->gpu.cudaCompCap < 80 && start->type != GPU) {
if (revLink == NULL) NCCLCHECK(findRevLink(node, link->remNode, &revLink));
revSpeed += fwSpeed/8;
revBw += fwBw/8;
}
if (link->remNode->type == CPU && link->type == LINK_NVL) {
if (revLink == NULL) NCCLCHECK(findRevLink(node, link->remNode, &revLink));
revSpeed += fwSpeed;
revBw += fwBw;
}
if (link->width < fwSpeed || (revSpeed && revLink->width < revSpeed)) { *steps = step; return ncclSuccess; }
SUB_ROUND(link->width, fwSpeed);
if (revSpeed) SUB_ROUND(revLink->width, revSpeed);
if (link->bw < fwBw || (revBw && revLink->bw < revBw)) { *steps = step; return ncclSuccess; }
SUB_ROUND(link->bw, fwBw);
if (revBw) SUB_ROUND(revLink->bw, revBw);
node = link->remNode;
}
*steps = maxSteps;
@@ -114,16 +114,16 @@ static ncclResult_t ncclTopoFollowPath(struct ncclTopoSystem* system, struct ncc
// Now check link type
*node = NULL;
int intra = type1 == GPU && type2 == GPU;
float speed = intra ? graph->speedIntra : graph->speedInter;
float bw = intra ? graph->bwIntra : graph->bwInter;
int type = intra ? graph->typeIntra : graph->typeInter;
if (mult == 1 && (path->type > type)) return ncclSuccess;
speed *= mult;
bw *= mult;
// Check there is enough bandwidth on paths.
int step = 0;
NCCLCHECK(followPath(path, node1, path->count, speed, &step));
NCCLCHECK(followPath(path, node1, path->count, bw, &step));
if (step < path->count) goto rewind;
// Enough bandwidth : return destination node.
@@ -133,11 +133,11 @@ static ncclResult_t ncclTopoFollowPath(struct ncclTopoSystem* system, struct ncc
rewind:
// Not enough bandwidth : rewind and exit.
NCCLCHECK(followPath(path, node1, step, -speed, &step));
NCCLCHECK(followPath(path, node1, step, -bw, &step));
return ncclSuccess;
}
static int gpuPciWidth(struct ncclTopoNode* gpu) {
static int gpuPciBw(struct ncclTopoNode* gpu) {
for (int l=0; l<gpu->nlinks; l++) {
struct ncclTopoLink* gpuLink = gpu->links+l;
if (gpuLink->type != LINK_PCI) continue;
@@ -145,7 +145,7 @@ static int gpuPciWidth(struct ncclTopoNode* gpu) {
for (int l=0; l<pci->nlinks; l++) {
struct ncclTopoLink* pciLink = pci->links+l;
if (pciLink->remNode != gpu) continue;
return std::min(gpuLink->width, pciLink->width);
return std::min(gpuLink->bw, pciLink->bw);
}
}
return -1;
@@ -157,29 +157,29 @@ struct ncclGpuScore {
int g; // Retain the index
int startIndex; // Least important
int intraNhops;
int intraWidth;
int intraBw;
int interNhops;
int interPciWidth;
int interWidth; // Most important
int interPciBw;
int interBw; // Most important
};
static int cmpScore(const void * g1, const void * g2) {
struct ncclGpuScore *s1 = (struct ncclGpuScore*)g1;
struct ncclGpuScore *s2 = (struct ncclGpuScore*)g2;
int d;
if ((d = (s2->interWidth - s1->interWidth))) return d;
if ((d = (s2->interPciWidth - s1->interPciWidth))) return d;
if ((d = (s2->interBw - s1->interBw))) return d;
if ((d = (s2->interPciBw - s1->interPciBw))) return d;
if ((d = (s1->interNhops - s2->interNhops))) return d;
if ((d = (s2->intraWidth - s1->intraWidth))) return d;
if ((d = (s2->intraBw - s1->intraBw))) return d;
if ((d = (s1->intraNhops - s2->intraNhops))) return d;
return s1->startIndex - s2->startIndex;
}
static int cmpIntraScores(struct ncclGpuScore* scores, int count) {
int intraWidth = scores[0].intraWidth;
int intraBw = scores[0].intraBw;
int intraNhops = scores[0].intraNhops;
for (int i=1; i<count; i++) {
if (scores[i].intraWidth != intraWidth || scores[i].intraNhops != intraNhops) return 1;
if (scores[i].intraBw != intraBw || scores[i].intraNhops != intraNhops) return 1;
}
return 0;
}
@@ -234,11 +234,11 @@ ncclResult_t ncclTopoSearchNextGpuSort(struct ncclTopoSystem* system, struct ncc
scores[count].g = g;
scores[count].startIndex = i;
scores[count].intraNhops = paths[g].count;
scores[count].intraWidth = paths[g].width;
scores[count].intraBw = paths[g].bw;
if (netPaths) {
scores[count].interNhops = netPaths[g].count;
scores[count].interPciWidth = gpuPciWidth(system->nodes[GPU].nodes+g);
scores[count].interWidth = netPaths[g].width;
scores[count].interPciBw = gpuPciBw(system->nodes[GPU].nodes+g);
scores[count].interBw = netPaths[g].bw;
}
count++;
}
@@ -341,8 +341,8 @@ ncclResult_t ncclTopoCompareGraphs(struct ncclTopoSystem* system, struct ncclTop
if (graph->nChannels < graph->minChannels) return ncclSuccess;
// 2. Try to get better bandwidth
if (graph->nChannels*graph->speedIntra < refGraph->nChannels*refGraph->speedIntra) return ncclSuccess;
if (graph->nChannels*graph->speedIntra > refGraph->nChannels*refGraph->speedIntra) {
if (graph->nChannels*graph->bwIntra < refGraph->nChannels*refGraph->bwIntra) return ncclSuccess;
if (graph->nChannels*graph->bwIntra > refGraph->nChannels*refGraph->bwIntra) {
*copy = 1;
return ncclSuccess;
}
@@ -446,23 +446,23 @@ ncclResult_t ncclTopoSearchRecGpu(struct ncclTopoSystem* system, struct ncclTopo
// Balanced Tree : count half of the bandwidth on first two GPUs
int nextBackToNet = -1;
float speedInterSave = graph->speedInter;
float bwInterSave = graph->bwInter;
if (graph->pattern == NCCL_TOPO_PATTERN_BALANCED_TREE) {
// Count half of the bandwidth on each of the first two GPUs
if (step == 0) nextBackToNet = 1;
else if (net->id != graph->inter[graph->nChannels*2+1]) continue;
graph->speedInter /= 2;
graph->bwInter /= 2;
}
NCCLCHECK(ncclTopoFollowPath(system, graph, GPU, g, NET, n, 1, &net));
graph->speedInter = speedInterSave;
graph->bwInter = bwInterSave;
if (net) {
graph->inter[graph->nChannels*2+1] = net->id;
NCCLCHECK(ncclTopoSearchRecGpu(system, graph, saveGraph, gpu, step, nextBackToNet, backToFirstRank, forcedOrder, time));
if (graph->pattern == NCCL_TOPO_PATTERN_BALANCED_TREE) graph->speedInter /= 2;
if (graph->pattern == NCCL_TOPO_PATTERN_BALANCED_TREE) graph->bwInter /= 2;
NCCLCHECK(ncclTopoFollowPath(system, graph, GPU, g, NET, n, -1, &net));
graph->speedInter = speedInterSave;
graph->bwInter = bwInterSave;
}
}
free(nets);
@@ -501,7 +501,7 @@ ncclResult_t ncclTopoSearchRecGpu(struct ncclTopoSystem* system, struct ncclTopo
}
ncclResult_t ncclTopoSearchRecNet(struct ncclTopoSystem* system, struct ncclTopoGraph* graph, struct ncclTopoGraph* saveGraph, int backToNet, int backToFirstRank, int* time) {
const int speed = graph->speedInter;
const int bw = graph->bwInter;
int* nets;
NCCLCHECK(ncclCalloc(&nets, system->nodes[NET].count));
int netcount;
@@ -511,7 +511,7 @@ ncclResult_t ncclTopoSearchRecNet(struct ncclTopoSystem* system, struct ncclTopo
struct ncclTopoNode* net = system->nodes[NET].nodes+n;
struct ncclTopoNode* gpu;
if (graph->collNet && net->net.collSupport == 0) continue;
if (net->net.width < speed) continue;
if (net->net.bw < bw) continue;
if (net->net.maxChannels == 0) continue;
graph->inter[graph->nChannels*2] = net->id;
@@ -520,7 +520,7 @@ ncclResult_t ncclTopoSearchRecNet(struct ncclTopoSystem* system, struct ncclTopo
for (int i=0; i<system->nodes[NET].count; i++) {
if ((system->nodes[NET].nodes[i].net.asic == net->net.asic) &&
(system->nodes[NET].nodes[i].net.port == net->net.port)) {
system->nodes[NET].nodes[i].net.width -= speed;
system->nodes[NET].nodes[i].net.bw -= bw;
}
}
net->net.maxChannels--;
@@ -554,26 +554,26 @@ ncclResult_t ncclTopoSearchRecNet(struct ncclTopoSystem* system, struct ncclTopo
}
// Then try the most local GPUs
float maxWidth = 0;
float maxBw = 0;
int minHops = 0xfffffff;
struct ncclTopoLinkList* paths = net->paths[GPU];
for (int g=0; g<system->nodes[GPU].count; g++) {
if (paths[g].width > maxWidth) {
maxWidth = paths[g].width;
if (paths[g].bw > maxBw) {
maxBw = paths[g].bw;
minHops = paths[g].count;
} else if (paths[g].width == maxWidth && paths[g].count < minHops) {
} else if (paths[g].bw == maxBw && paths[g].count < minHops) {
minHops = paths[g].count;
}
}
if (maxWidth >= speed) {
if (maxBw >= bw) {
// In the first loop, avoid using GPUs in both directions between channels (one channel
// sending from that GPU and one channel receiving to that GPU), since that usually leads
// to lower BW.
for (int tryGpuBidir=0; tryGpuBidir<2; tryGpuBidir++) {
for (int g=0; g<system->nodes[GPU].count; g++) {
if (paths[g].width == maxWidth && paths[g].count == minHops) {
if (paths[g].bw == maxBw && paths[g].count == minHops) {
gpu = system->nodes[GPU].nodes+g;
int gpuUsed = gpuPciWidth(gpu) > 0 ? 0 : 1;
int gpuUsed = gpuPciBw(gpu) > 0 ? 0 : 1;
if (tryGpuBidir == gpuUsed) {
NCCLCHECK(ncclTopoSearchTryGpu(system, graph, saveGraph, 0, backToNet, backToFirstRank, 0, time, NET, n, g));
}
@@ -587,7 +587,7 @@ ncclResult_t ncclTopoSearchRecNet(struct ncclTopoSystem* system, struct ncclTopo
for (int i=0; i<system->nodes[NET].count; i++) {
if ((system->nodes[NET].nodes[i].net.asic == net->net.asic) &&
(system->nodes[NET].nodes[i].net.port == net->net.port)) {
system->nodes[NET].nodes[i].net.width += speed;
system->nodes[NET].nodes[i].net.bw += bw;
}
}
}
@@ -706,8 +706,8 @@ ncclResult_t ncclTopoGetGraphFromXmlSub(struct ncclXmlNode *xmlGraph, struct ncc
NCCLCHECK(xmlGetAttrInt(xmlGraph, "pattern", &graph->pattern));
NCCLCHECK(xmlGetAttrInt(xmlGraph, "nchannels", &graph->nChannels));
NCCLCHECK(xmlGetAttrFloat(xmlGraph, "speedintra", &graph->speedIntra));
NCCLCHECK(xmlGetAttrFloat(xmlGraph, "speedinter", &graph->speedInter));
NCCLCHECK(xmlGetAttrFloat(xmlGraph, "speedintra", &graph->bwIntra));
NCCLCHECK(xmlGetAttrFloat(xmlGraph, "speedinter", &graph->bwInter));
if (xmlGetAttrFloat(xmlGraph, "latencyinter", &graph->latencyInter) != ncclSuccess) graph->latencyInter = 0.0;
const char* str;
NCCLCHECK(xmlGetAttr(xmlGraph, "typeintra", &str));
@@ -767,8 +767,8 @@ ncclResult_t ncclTopoGetXmlFromGraph(struct ncclTopoGraph* graph, struct ncclTop
NCCLCHECK(xmlSetAttrInt(xmlGraph, "pattern", graph->pattern));
NCCLCHECK(xmlSetAttrInt(xmlGraph, "crossnic", graph->crossNic));
NCCLCHECK(xmlSetAttrInt(xmlGraph, "nchannels", graph->nChannels));
NCCLCHECK(xmlSetAttrFloat(xmlGraph, "speedintra", graph->speedIntra));
NCCLCHECK(xmlSetAttrFloat(xmlGraph, "speedinter", graph->speedInter));
NCCLCHECK(xmlSetAttrFloat(xmlGraph, "speedintra", graph->bwIntra));
NCCLCHECK(xmlSetAttrFloat(xmlGraph, "speedinter", graph->bwInter));
NCCLCHECK(xmlSetAttrFloat(xmlGraph, "latencyinter", graph->latencyInter));
const char* str;
NCCLCHECK(kvConvertToStr(graph->typeIntra, &str, kvDictLinkType));
@@ -830,7 +830,7 @@ ncclResult_t ncclTopoCompute(ncclTopoSystem* system, struct ncclTopoGraph* graph
int ngpus = system->nodes[GPU].count;
graph->crossNic = ncclParamCrossNic();
int crossNic = (system->nodes[NET].count > 1) && graph->crossNic ? 1 : 0;
graph->speedIntra = graph->speedInter = 0;
graph->bwIntra = graph->bwInter = 0;
graph->latencyInter = 0;
if (graph->crossNic == 2) graph->crossNic = 0;
graph->typeIntra = ngpus == 1 ? PATH_LOC : PATH_NVL;
@@ -903,7 +903,7 @@ ncclResult_t ncclTopoCompute(ncclTopoSystem* system, struct ncclTopoGraph* graph
struct ncclTopoGraph tmpGraph;
memcpy(&tmpGraph, graph, sizeof(struct ncclTopoGraph));
// First try crossnic, then decrease speed and finally increase speedIntra.
// First try crossnic, then decrease bw and finally increase bwIntra.
int nspeeds = 0;
float* speedArray = NULL;
if (system->nodes[NET].count == 0) {
@@ -915,8 +915,8 @@ ncclResult_t ncclTopoCompute(ncclTopoSystem* system, struct ncclTopoGraph* graph
}
int pass = 1;
int speedIndex = 0;
while (speedArray[speedIndex] > system->maxWidth && speedIndex < nspeeds-1) speedIndex++;
tmpGraph.speedIntra = tmpGraph.speedInter = speedArray[speedIndex];
while (speedArray[speedIndex] > system->maxBw && speedIndex < nspeeds-1) speedIndex++;
tmpGraph.bwIntra = tmpGraph.bwInter = speedArray[speedIndex];
int64_t globalTimeout = NCCL_SEARCH_GLOBAL_TIMEOUT;
search:
int time = tmpGraph.sameChannels ? NCCL_SEARCH_TIMEOUT_SAMECHANNELS :
@@ -926,7 +926,7 @@ search:
NCCLCHECK(ncclTopoSearchRec(system, &tmpGraph, graph, &time));
#if 0
printf("Pattern %d, crossNic %d, Speed %g/%g, type %d/%d, channels %d-%d sameChannels %d -> nChannels %dx%g/%g %s\n", tmpGraph.pattern, tmpGraph.crossNic, tmpGraph.speedInter, tmpGraph.speedIntra, tmpGraph.typeInter, tmpGraph.typeIntra, tmpGraph.minChannels, tmpGraph.maxChannels, tmpGraph.sameChannels, graph->nChannels, graph->speedInter, graph->speedIntra, time == 0 ? "TIMEOUT" : time == -1 ? "PERFECT" : "");
printf("Pattern %d, crossNic %d, Bw %g/%g, type %d/%d, channels %d-%d sameChannels %d -> nChannels %dx%g/%g %s\n", tmpGraph.pattern, tmpGraph.crossNic, tmpGraph.bwInter, tmpGraph.bwIntra, tmpGraph.typeInter, tmpGraph.typeIntra, tmpGraph.minChannels, tmpGraph.maxChannels, tmpGraph.sameChannels, graph->nChannels, graph->bwInter, graph->bwIntra, time == 0 ? "TIMEOUT" : time == -1 ? "PERFECT" : "");
for (int c=0; c<graph->nChannels; c++) {
printf("%2d : ", c);
for (int g=0; g<ngpus; g++) {
@@ -938,7 +938,7 @@ search:
#endif
// Optimal solution, stop here
if (time == -1) goto done;
if (graph->nChannels*graph->speedInter >= system->totalWidth) goto done;
if (graph->nChannels*graph->bwInter >= system->totalBw) goto done;
if (pass == 1) {
// First pass, we don't have a solution yet ; try other options
@@ -980,14 +980,14 @@ search:
}
tmpGraph.pattern = graph->pattern;
// Decrease speed until we find a solution
if ((speedIndex < nspeeds-1) && (graph->nChannels == 0 || (speedArray[speedIndex+1]/graph->speedInter > .49))) {
tmpGraph.speedInter = tmpGraph.speedIntra = speedArray[++speedIndex];
// Decrease bw until we find a solution
if ((speedIndex < nspeeds-1) && (graph->nChannels == 0 || (speedArray[speedIndex+1]/graph->bwInter > .49))) {
tmpGraph.bwInter = tmpGraph.bwIntra = speedArray[++speedIndex];
goto search;
}
speedIndex = 0;
while (speedArray[speedIndex] > system->maxWidth && speedIndex < nspeeds-1) speedIndex++;
tmpGraph.speedIntra = tmpGraph.speedInter = speedArray[speedIndex];
while (speedArray[speedIndex] > system->maxBw && speedIndex < nspeeds-1) speedIndex++;
tmpGraph.bwIntra = tmpGraph.bwInter = speedArray[speedIndex];
}
@@ -997,18 +997,18 @@ done:
time = -1;
memcpy(&tmpGraph, graph, sizeof(tmpGraph));
speedIndex = 0;
while (speedArray[speedIndex] > graph->speedInter && speedIndex < nspeeds-1) speedIndex++;
tmpGraph.speedIntra = tmpGraph.speedInter = speedArray[speedIndex];
while (speedArray[speedIndex] > graph->bwInter && speedIndex < nspeeds-1) speedIndex++;
tmpGraph.bwIntra = tmpGraph.bwInter = speedArray[speedIndex];
tmpGraph.minChannels = graph->nChannels;
pass = 2;
}
// 3. See if we can increase speedIntra for trees (2 nodes or collnet)
// 3. See if we can increase bwIntra for trees (2 nodes or collnet)
if (pass == 2) {
if (time != 0 && graph->pattern != NCCL_TOPO_PATTERN_RING &&
tmpGraph.speedIntra == graph->speedIntra && tmpGraph.speedIntra < tmpGraph.speedInter*2 &&
tmpGraph.bwIntra == graph->bwIntra && tmpGraph.bwIntra < tmpGraph.bwInter*2 &&
speedIndex > 0) {
tmpGraph.speedIntra = speedArray[--speedIndex];
tmpGraph.bwIntra = speedArray[--speedIndex];
goto search;
}
time = -1;
@@ -1019,17 +1019,17 @@ done:
WARN("Could not find a path for pattern %d, falling back to simple order", graph->pattern);
for (int i=0; i<ngpus; i++) graph->intra[i] = system->nodes[GPU].nodes[i].gpu.rank[0];
graph->inter[0] = graph->inter[1] = 0;
graph->speedIntra = graph->speedInter = 0.1;
graph->bwIntra = graph->bwInter = 0.1;
graph->typeIntra = graph->typeInter = PATH_SYS;
graph->nChannels = 1;
}
if (graph->speedIntra >= 25.0) {
if (graph->bwIntra >= 25.0) {
int dupChannels = std::min(graph->nChannels*2, graph->maxChannels);
memcpy(graph->intra+graph->nChannels*ngpus, graph->intra, (dupChannels-graph->nChannels)*ngpus*sizeof(int));
memcpy(graph->inter+graph->nChannels*2,graph->inter, (dupChannels-graph->nChannels)*2*sizeof(int));
memcpy(graph->intraNets+graph->nChannels*ngpus*2, graph->intraNets, (dupChannels-graph->nChannels)*2*ngpus*sizeof(int));
graph->speedIntra /= DIVUP(dupChannels, graph->nChannels);
graph->speedInter /= DIVUP(dupChannels, graph->nChannels);
graph->bwIntra /= DIVUP(dupChannels, graph->nChannels);
graph->bwInter /= DIVUP(dupChannels, graph->nChannels);
graph->nChannels = dupChannels;
}
ncclExpandMultiRank(system, graph);
@@ -1037,7 +1037,7 @@ done:
}
ncclResult_t ncclTopoPrintGraph(struct ncclTopoSystem* system, struct ncclTopoGraph* graph) {
INFO(NCCL_GRAPH, "Pattern %d, crossNic %d, nChannels %d, speed %f/%f, type %s/%s, sameChannels %d", graph->pattern, graph->crossNic, graph->nChannels, graph->speedIntra, graph->speedInter, topoPathTypeStr[graph->typeIntra], topoPathTypeStr[graph->typeInter], graph->sameChannels);
INFO(NCCL_GRAPH, "Pattern %d, crossNic %d, nChannels %d, bw %f/%f, type %s/%s, sameChannels %d", graph->pattern, graph->crossNic, graph->nChannels, graph->bwIntra, graph->bwInter, topoPathTypeStr[graph->typeIntra], topoPathTypeStr[graph->typeInter], graph->sameChannels);
int ngpus = system->nodes[GPU].count;
char line[1024];
src/graph/topo.cc
+37 -32
View File
@@ -62,24 +62,24 @@ static ncclResult_t findLocalCpu(struct ncclTopoNode* node, struct ncclTopoNode*
return ncclSuccess;
}
int interCpuWidth = 0;
int cpuPciWidth = 0;
int interCpuBw = 0;
int cpuPciBw = 0;
static ncclResult_t ncclTopoGetInterCpuWidth(struct ncclTopoNode* cpu, float* width) {
*width = LOC_WIDTH;
static ncclResult_t ncclTopoGetInterCpuBw(struct ncclTopoNode* cpu, float* bw) {
*bw = LOC_BW;
if (cpu->cpu.arch == NCCL_TOPO_CPU_ARCH_POWER) {
*width = P9_WIDTH;
*bw = P9_BW;
return ncclSuccess;
}
if (cpu->cpu.arch == NCCL_TOPO_CPU_ARCH_ARM) {
*width = ARM_WIDTH;
*bw = ARM_BW;
return ncclSuccess;
}
if (cpu->cpu.arch == NCCL_TOPO_CPU_ARCH_X86 && cpu->cpu.vendor == NCCL_TOPO_CPU_VENDOR_INTEL) {
*width = cpu->cpu.model == NCCL_TOPO_CPU_TYPE_SKL ? SKL_QPI_WIDTH : QPI_WIDTH;
*bw = cpu->cpu.model == NCCL_TOPO_CPU_TYPE_SKL ? SKL_QPI_BW : QPI_BW;
}
if (cpu->cpu.arch == NCCL_TOPO_CPU_ARCH_X86 && cpu->cpu.vendor == NCCL_TOPO_CPU_VENDOR_ZHAOXIN) {
*width = cpu->cpu.model == NCCL_TOPO_CPU_TYPE_YONGFENG ? YONGFENG_ZPI_WIDTH : ZPI_WIDTH;
*bw = cpu->cpu.model == NCCL_TOPO_CPU_TYPE_YONGFENG ? YONGFENG_ZPI_BW : ZPI_BW;
}
return ncclSuccess;
}
@@ -115,7 +115,7 @@ ncclResult_t ncclTopoCreateNode(struct ncclTopoSystem* system, struct ncclTopoNo
n->nlinks=1;
n->links[0].type = LINK_LOC;
n->links[0].remNode = n;
n->links[0].width = LOC_WIDTH;
n->links[0].bw = LOC_BW;
n->gpu.dev = NCCL_TOPO_UNDEF;
for (int i=0; i<RCCL_TOPO_MAX_RANKS_PER_GPU; i++) {
n->gpu.rank[i] = NCCL_TOPO_UNDEF;
@@ -129,7 +129,7 @@ ncclResult_t ncclTopoCreateNode(struct ncclTopoSystem* system, struct ncclTopoNo
} else if (type == NET) {
n->net.asic = 0ULL;
n->net.port = NCCL_TOPO_UNDEF;
n->net.width = 0.0;
n->net.bw = 0.0;
n->net.latency = 0.0;
}
*node = n;
@@ -159,8 +159,8 @@ ncclResult_t ncclTopoRemoveNode(struct ncclTopoSystem* system, int type, int ind
return ncclSuccess;
}
ncclResult_t ncclTopoConnectNodes(struct ncclTopoNode* node, struct ncclTopoNode* remNode, int type, float width) {
// Aggregate links into higher width for NVLink
ncclResult_t ncclTopoConnectNodes(struct ncclTopoNode* node, struct ncclTopoNode* remNode, int type, float bw) {
// Aggregate links into higher bw for NVLink
struct ncclTopoLink* link;
for (link = node->links; link->remNode; link++) {
if (link->remNode == remNode && link->type == type) break;
@@ -168,13 +168,13 @@ ncclResult_t ncclTopoConnectNodes(struct ncclTopoNode* node, struct ncclTopoNode
if (link->remNode == NULL) node->nlinks++;
link->type = type;
link->remNode = remNode;
link->width += width;
link->bw += bw;
// Sort links in BW descending order
struct ncclTopoLink linkSave;
memcpy(&linkSave, link, sizeof(struct ncclTopoLink));
while (link != node->links) {
if ((link-1)->width >= linkSave.width) break;
if ((link-1)->bw >= linkSave.bw) break;
memcpy(link, link-1, sizeof(struct ncclTopoLink));
link--;
}
@@ -246,9 +246,9 @@ ncclResult_t ncclTopoConnectCpus(struct ncclTopoSystem* system) {
for (int n=0; n<system->nodes[CPU].count; n++) {
for (int p=0; p<system->nodes[CPU].count; p++) {
if (n == p) continue;
float width;
NCCLCHECK(ncclTopoGetInterCpuWidth(system->nodes[CPU].nodes+n, &width));
NCCLCHECK(ncclTopoConnectNodes(system->nodes[CPU].nodes+n, system->nodes[CPU].nodes+p, LINK_SYS, width));
float bw;
NCCLCHECK(ncclTopoGetInterCpuBw(system->nodes[CPU].nodes+n, &bw));
NCCLCHECK(ncclTopoConnectNodes(system->nodes[CPU].nodes+n, system->nodes[CPU].nodes+p, LINK_SYS, bw));
}
}
return ncclSuccess;
@@ -279,13 +279,13 @@ static ncclResult_t ncclTopoPrintRec(struct ncclTopoNode* node, struct ncclTopoN
struct ncclTopoLink* link = node->links+l;
if (link->type == LINK_LOC) continue;
if (link->type != LINK_PCI || link->remNode != prevNode) {
sprintf(line+offset, "+ %s[%2.1f] - ", topoLinkTypeStr[link->type], link->width);
sprintf(line+offset, "+ %s[%2.1f] - ", topoLinkTypeStr[link->type], link->bw);
int nextOffset = strlen(line);
if (link->type == LINK_PCI) {
NCCLCHECK(ncclTopoPrintRec(link->remNode, node, line, nextOffset));
} else {
if (link->remNode->type == NET) {
sprintf(line+nextOffset, "%s/%lX (%lx/%d/%f)", topoNodeTypeStr[link->remNode->type], link->remNode->id, link->remNode->net.asic, link->remNode->net.port, link->remNode->net.width);
sprintf(line+nextOffset, "%s/%lX (%lx/%d/%f)", topoNodeTypeStr[link->remNode->type], link->remNode->id, link->remNode->net.asic, link->remNode->net.port, link->remNode->net.bw);
} else {
sprintf(line+nextOffset, "%s/%lX", topoNodeTypeStr[link->remNode->type], link->remNode->id);
}
@@ -297,7 +297,7 @@ static ncclResult_t ncclTopoPrintRec(struct ncclTopoNode* node, struct ncclTopoN
}
ncclResult_t ncclTopoPrint(struct ncclTopoSystem* s) {
INFO(NCCL_GRAPH, "=== System : maxWidth %2.1f totalWidth %2.1f ===", s->maxWidth, s->totalWidth);
INFO(NCCL_GRAPH, "=== System : maxBw %2.1f totalBw %2.1f ===", s->maxBw, s->totalBw);
char line[1024];
for (int n=0; n<s->nodes[CPU].count; n++) NCCLCHECK(ncclTopoPrintRec(s->nodes[CPU].nodes+n, NULL, line, 0));
INFO(NCCL_GRAPH, "==========================================");
@@ -352,7 +352,7 @@ ncclResult_t ncclTopoAddNet(struct ncclXmlNode* xmlNet, struct ncclTopoSystem* s
int mbps;
NCCLCHECK(xmlGetAttrIntDefault(xmlNet, "speed", &mbps, 0));
if (mbps <= 0) mbps = 10000; // Some NICs define speed = -1
net->net.width = mbps / 8000.0;
net->net.bw = mbps / 8000.0;
if (xmlGetAttrFloat(xmlNet, "latency", &net->net.latency) != ncclSuccess) net->net.latency = 0;
NCCLCHECK(xmlGetAttrIntDefault(xmlNet, "port", &net->net.port, 0));
NCCLCHECK(xmlGetAttrIntDefault(xmlNet, "gdr", &net->net.gdrSupport, 0));
@@ -361,8 +361,8 @@ ncclResult_t ncclTopoAddNet(struct ncclXmlNode* xmlNet, struct ncclTopoSystem* s
net->net.busId = busId;
ncclDebugNoWarn = 0;
NCCLCHECK(ncclTopoConnectNodes(nic, net, LINK_NET, net->net.width));
NCCLCHECK(ncclTopoConnectNodes(net, nic, LINK_NET, net->net.width));
NCCLCHECK(ncclTopoConnectNodes(nic, net, LINK_NET, net->net.bw));
NCCLCHECK(ncclTopoConnectNodes(net, nic, LINK_NET, net->net.bw));
return ncclSuccess;
}
@@ -520,8 +520,8 @@ ncclResult_t ncclTopoAddCpu(struct ncclXmlNode* xmlCpu, struct ncclTopoSystem* s
NCCLCHECK(ncclTopoGetNode(system, &nic, NIC, 0));
if (nic == NULL) {
NCCLCHECK(ncclTopoCreateNode(system, &nic, NIC, 0));
NCCLCHECK(ncclTopoConnectNodes(cpu, nic, LINK_PCI, LOC_WIDTH));
NCCLCHECK(ncclTopoConnectNodes(nic, cpu, LINK_PCI, LOC_WIDTH));
NCCLCHECK(ncclTopoConnectNodes(cpu, nic, LINK_PCI, LOC_BW));
NCCLCHECK(ncclTopoConnectNodes(nic, cpu, LINK_PCI, LOC_BW));
}
NCCLCHECK(ncclTopoAddNic(node, system, nic, 0));
}
@@ -616,10 +616,10 @@ ncclResult_t ncclTopoAddNvLinks(struct ncclXmlNode* node, struct ncclTopoSystem*
}
}
if (remote) {
float nvlSpeed = ncclTopoNVLinkSpeed(gpu->gpu.cudaCompCap);
NCCLCHECK(ncclTopoConnectNodes(gpu, remote, LINK_NVL, count*nvlSpeed));
float nvlBw = ncclTopoNVLinkBw(gpu->gpu.cudaCompCap);
NCCLCHECK(ncclTopoConnectNodes(gpu, remote, LINK_NVL, count*nvlBw));
if (remote->type != GPU) {
NCCLCHECK(ncclTopoConnectNodes(remote, gpu, LINK_NVL, count*nvlSpeed));
NCCLCHECK(ncclTopoConnectNodes(remote, gpu, LINK_NVL, count*nvlBw));
}
}
} else {
@@ -781,18 +781,18 @@ ncclResult_t ncclTopoGetLocalNet(struct ncclTopoSystem* system, int rank, int* i
int g;
NCCLCHECK(ncclTopoRankToIndex(system, rank, &g));
int minType = PATH_SYS;
float maxWidth = 0;
float maxBw = 0;
int count = 0;
int* nets;
NCCLCHECK(ncclCalloc(&nets, system->nodes[NET].count));
for (int n=0; n<system->nodes[NET].count; n++) {
struct ncclTopoLinkList* path = system->nodes[NET].nodes[n].paths[GPU]+g;
if (path->width > maxWidth || (path->width == maxWidth && path->type < minType)) {
maxWidth = path->width;
if (path->bw > maxBw || (path->bw == maxBw && path->type < minType)) {
maxBw = path->bw;
minType = path->type;
count = 0;
}
if (path->width == maxWidth && path->type == minType) nets[count++] = system->nodes[NET].nodes[n].id;
if (path->bw == maxBw && path->type == minType) nets[count++] = system->nodes[NET].nodes[n].id;
}
if (count == 0) {
*id = -1;
@@ -890,6 +890,11 @@ ncclResult_t ncclTopoGetNetCount(struct ncclTopoSystem* system, int* count) {
return ncclSuccess;
}
ncclResult_t ncclTopoGetNvsCount(struct ncclTopoSystem* system, int* count) {
*count = system->nodes[NVS].count;
return ncclSuccess;
}
ncclResult_t ncclTopoGetCompCap(struct ncclTopoSystem* system, int* ccMin, int* ccMax) {
if (system->nodes[GPU].count == 0) return ncclInternalError;
int min, max;
src/graph/topo.h
+20 -20
View File
@@ -11,25 +11,25 @@
#include "graph.h"
#include "core.h"
#define LOC_WIDTH 5000.0
#define SM60_NVLINK_WIDTH 18.0
#define SM70_NVLINK_WIDTH 22.0
#define SM80_NVLINK_WIDTH 22.0
#define SM86_NVLINK_WIDTH 12.0
#define PCI_WIDTH 12.0 // PCI Gen3 x16
#define QPI_WIDTH 6.0
#define SKL_QPI_WIDTH 9.0
#define ZPI_WIDTH 6.0
#define YONGFENG_ZPI_WIDTH 9.0
#define P9_WIDTH 32.0
#define ARM_WIDTH 6.0
#define NET_WIDTH 12.0 // 100Gbit
#define LOC_BW 5000.0
#define SM60_NVLINK_BW 18.0
#define SM70_NVLINK_BW 22.0
#define SM80_NVLINK_BW 22.0
#define SM86_NVLINK_BW 12.0
#define PCI_BW 12.0 // PCI Gen3 x16
#define QPI_BW 6.0
#define SKL_QPI_BW 9.0
#define ZPI_BW 6.0
#define YONGFENG_ZPI_BW 9.0
#define P9_BW 32.0
#define ARM_BW 6.0
#define NET_BW 12.0 // 100Gbit
#define VEGA_XGMI_WIDTH 24.0
#define MI200_XGMI_WIDTH 36.0
// Intel CPU convert GPU P2P traffic into 64B PCI TLPs, so GPU
// to GPU traffic consumes more PCI bandwidth.
#define INTEL_P2P_OVERHEAD(speed) (speed*6/5)
#define INTEL_P2P_OVERHEAD(bw) (bw*6/5)
#define NCCL_TOPO_NODE_TYPES 7
#define GPU 0
@@ -81,7 +81,7 @@ extern const char* topoPathTypeStr[];
struct ncclTopoNode;
struct ncclTopoLink {
int type;
float width;
float bw;
struct ncclTopoNode* remNode;
};
#define NCCL_TOPO_MAX_LINKS 32
@@ -90,7 +90,7 @@ struct ncclTopoLink {
struct ncclTopoLinkList {
struct ncclTopoLink* list[NCCL_TOPO_MAX_HOPS];
int count;
float width;
float bw;
int type;
};
@@ -124,7 +124,7 @@ struct ncclTopoNode {
struct {
uint64_t asic;
int port;
float width;
float bw;
float latency;
int gdrSupport;
int collSupport;
@@ -156,8 +156,8 @@ struct ncclTopoNodeSet {
struct ncclTopoSystem {
struct ncclTopoNodeSet nodes[NCCL_TOPO_NODE_TYPES];
float maxWidth;
float totalWidth;
float maxBw;
float totalBw;
int type;
int nRanks;
int netGdrLevel;
@@ -172,7 +172,7 @@ struct ncclTopoSystem {
ncclResult_t ncclTopoGetNode(struct ncclTopoSystem* system, struct ncclTopoNode** node, int type, uint64_t id);
ncclResult_t ncclTopoCreateNode(struct ncclTopoSystem* system, struct ncclTopoNode** node, int type, uint64_t id);
ncclResult_t ncclTopoRemoveNode(struct ncclTopoSystem* system, int type, int id);
ncclResult_t ncclTopoConnectNodes(struct ncclTopoNode* node, struct ncclTopoNode* remNode, int type, float width);
ncclResult_t ncclTopoConnectNodes(struct ncclTopoNode* node, struct ncclTopoNode* remNode, int type, float bw);
ncclResult_t ncclTopoPrintPaths(struct ncclTopoSystem* system);
ncclResult_t ncclTopoLoadSystem(const char* xmlTopoFile, struct ncclTopoSystem* system);
ncclResult_t ncclTopoGetIntermediateRank(struct ncclTopoSystem* system, int rank, int netDev, int* intermediateRank);
src/graph/tuning.cc
+66 -46
View File
@@ -71,18 +71,18 @@ struct tuningModel {
static struct tuningModel tuning_model_0 {
.hwLat = {
/* NVLINK */
{ /* Tree (LL/LL128/Simple)*/ { 1.5, 1.5, 4.5 }, /* Ring (LL/LL128/Simple)*/ { 1.5, 1.5, 4.5 }, /* CollNet (LL/LL128/Simple)*/ { 1.5, 1.5, 4.5 } },
{ /* Tree (LL/LL128/Simple)*/ { 1.5, 1.5, 4.5 }, /* Ring (LL/LL128/Simple)*/ { 1.5, 1.5, 4.5 }, /* CollNetDirect (Simple)*/ { 0.0, 0.0, 4.5 }, /* CollNetChain (Simple)*/ { 0.0, 0.0, 4.5 } },
/* PCI */
{ /* Tree (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 }, /* Ring (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 }, /* CollNet (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 } },
{ /* Tree (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 }, /* Ring (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 }, /* CollNetDirect (Simple)*/ { 0.0, 0.0, 5.7 }, /* CollNetChain (Simple)*/ { 0.0, 0.0, 5.7 } },
/* NET */
{ /* Tree (LL/LL128/Simple)*/ { 28.3, 28.3, 45.4 }, /* Ring (LL/LL128/Simple)*/ { 2.0, 2.0, 24.1 }, /* CollNet (LL/LL128/Simple)*/ { 28.3, 28.3, 45.4 } },
{ /* Tree (LL/LL128/Simple)*/ { 28.3, 28.3, 45.4 }, /* Ring (LL/LL128/Simple)*/ { 2.0, 2.0, 24.1 }, /* CollNetDirect (Simple)*/ { 0.0, 0.0, 45.4 }, /* CollNetChain (Simple)*/ { 0.0, 0.0, 45.4 } },
},
.bwRatio = {
/* 2 nodes */
{ /* Tree (LL/LL128/Simple)*/ { 0.06, 1.00, 1.30 }, /* Ring (LL/LL128/Simple)*/ { 0.07, 1.00, 1.00 }, /* CollNet (LL/LL128/Simple)*/ { 1.00, 1.00, 1.00 } },
{ /* Tree (LL/LL128/Simple)*/ { 0.06, 1.00, 1.30 }, /* Ring (LL/LL128/Simple)*/ { 0.07, 1.00, 1.00 }, /* CollNetDirect (Simple)*/ { 0.00, 0.00, 1.00 }, /* CollNetChain (Simple)*/ { 0.00, 0.00, 1.00 } },
/* more than 2 nodes */
{ /* Tree (LL/LL128/Simple)*/ { 0.06, 1.00, 0.30 }, /* Ring (LL/LL128/Simple)*/ { 0.07, 1.00, 1.00 }, /* CollNet (LL/LL128/Simple)*/ { 1.00, 1.00, 1.00 } },
{ /* Tree (LL/LL128/Simple)*/ { 0.06, 1.00, 0.30 }, /* Ring (LL/LL128/Simple)*/ { 0.07, 1.00, 1.00 }, /* CollNetDirect (Simple)*/ { 0.00, 0.00, 1.00 }, /* CollNetChain (Simple)*/ { 0.00, 0.00, 1.00 } },
},
.treeCorrectionFactor = {
@@ -101,18 +101,18 @@ static struct tuningModel tuning_model_0 {
static struct tuningModel tuning_model_1 {
.hwLat =
{ /* NVLINK */
{ /* Tree (LL/LL128/Simple)*/ { 1.5, 1.5, 4.5 }, /* Ring (LL/LL128/Simple)*/ { 1.5, 1.5, 4.5 }, /* CollNet (LL/LL128/Simple)*/ { 1.5, 1.5, 4.5 } },
{ /* Tree (LL/LL128/Simple)*/ { 1.5, 1.5, 4.5 }, /* Ring (LL/LL128/Simple)*/ { 1.5, 1.5, 4.5 }, /* CollNetDirect (Simple)*/ { 0.0, 0.0, 4.5 }, /* CollNetChain (Simple)*/ { 0.0, 0.0, 4.5 } },
/* PCI */
{ /* Tree (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 }, /* Ring (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 }, /* CollNet (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 } },
{ /* Tree (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 }, /* Ring (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 }, /* CollNetDirect (Simple)*/ { 0.0, 0.0, 5.7 }, /* CollNetChain (Simple)*/ { 0.0, 0.0, 5.7 } },
/* NET */
{ /* Tree (LL/LL128/Simple)*/ { 33.0, 33.0, 15.8 }, /* Ring (LL/LL128/Simple)*/ { 5.1, 5.1, 68.8 }, /* CollNet (LL/LL128/Simple)*/ { 33.0, 33.0, 15.8 } },
{ /* Tree (LL/LL128/Simple)*/ { 33.0, 33.0, 15.8 }, /* Ring (LL/LL128/Simple)*/ { 5.1, 5.1, 68.8 }, /* CollNetDirect (Simple)*/ { 0.0, 0.0, 15.8 }, /* CollNetChain (Simple)*/ { 0.0, 0.0, 15.8 } },
},
.bwRatio =
{ /* 2 nodes */
{ /* Tree (LL/LL128/Simple)*/ { 0.12, 1.00, 0.99 }, /* Ring (LL/LL128/Simple)*/ { 0.12, 1.00, 1.00 }, /* CollNet (LL/LL128/Simple)*/ { 1.00, 1.00, 1.00 } },
{ /* Tree (LL/LL128/Simple)*/ { 0.12, 1.00, 0.99 }, /* Ring (LL/LL128/Simple)*/ { 0.12, 1.00, 1.00 }, /* CollNetDirect (Simple)*/ { 0.00, 0.00, 1.00 }, /* CollNetChain (Simple)*/ { 0.00, 0.00, 1.00 } },
/* more than 2 nodes */
{ /* Tree (LL/LL128/Simple)*/ { 0.15, 1.00, 0.42 }, /* Ring (LL/LL128/Simple)*/ { 0.20, 1.00, 1.00 }, /* CollNet (LL/LL128/Simple)*/ { 1.00, 1.00, 1.00 } },
{ /* Tree (LL/LL128/Simple)*/ { 0.15, 1.00, 0.42 }, /* Ring (LL/LL128/Simple)*/ { 0.20, 1.00, 1.00 }, /* CollNetDirect (Simple)*/ { 0.00, 0.00, 1.00 }, /* CollNetChain (Simple)*/ { 0.00, 0.00, 1.00 } },
},
.treeCorrectionFactor = {
@@ -131,18 +131,18 @@ static struct tuningModel tuning_model_1 {
static struct tuningModel tuning_model_2 {
.hwLat = {
/* NVLINK */
{ /* Tree (LL/LL128/Simple)*/ { 1.5, 1.5, 4.5 }, /* Ring (LL/LL128/Simple)*/ { 1.5, 1.5, 4.5 }, /* CollNet (LL/LL128/Simple)*/ { 1.5, 1.5, 4.5 } },
{ /* Tree (LL/LL128/Simple)*/ { 1.5, 1.5, 4.5 }, /* Ring (LL/LL128/Simple)*/ { 1.5, 1.5, 4.5 }, /* CollNetDirect (Simple)*/ { 0.0, 0.0, 4.5 }, /* CollNetChain (Simple)*/ { 0.0, 0.0, 4.5 } },
/* PCI */
{ /* Tree (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 }, /* Ring (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 }, /* CollNet (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 } },
{ /* Tree (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 }, /* Ring (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 }, /* CollNetDirect (Simple)*/ { 0.0, 0.0, 5.7 }, /* CollNetChain (Simple)*/ { 0.0, 0.0, 5.7 } },
/* NET */
{ /* Tree (LL/LL128/Simple)*/ { 27.9, 27.9, 15.8 }, /* Ring (LL/LL128/Simple)*/ { 12.1, 12.1, 68.8 }, /* CollNet (LL/LL128/Simple)*/ { 27.9, 27.9, 15.8 } },
{ /* Tree (LL/LL128/Simple)*/ { 27.9, 27.9, 15.8 }, /* Ring (LL/LL128/Simple)*/ { 12.1, 12.1, 68.8 }, /* CollNetDirect (Simple)*/ { 0.0, 0.0, 15.8 }, /* CollNetChain (Simple)*/ { 0.0, 0.0, 15.8 } },
},
.bwRatio = {
/* 2 nodes */
{ /* Tree (LL/LL128/Simple)*/ { 0.07, 1.00, 0.99 }, /* Ring (LL/LL128/Simple)*/ { 0.08, 1.00, 1.00 }, /* CollNet (LL/LL128/Simple)*/ { 1.00, 1.00, 1.00 } },
{ /* Tree (LL/LL128/Simple)*/ { 0.07, 1.00, 0.99 }, /* Ring (LL/LL128/Simple)*/ { 0.08, 1.00, 1.00 }, /* CollNetDirect (Simple)*/ { 0.00, 0.00, 1.00 }, /* CollNetChain (Simple)*/ { 0.00, 0.00, 1.00 } },
/* more than 2 nodes */
{ /* Tree (LL/LL128/Simple)*/ { 0.07, 1.00, 0.42 }, /* Ring (LL/LL128/Simple)*/ { 0.08, 1.00, 1.00 }, /* CollNet (LL/LL128/Simple)*/ { 1.00, 1.00, 1.00 } },
{ /* Tree (LL/LL128/Simple)*/ { 0.07, 1.00, 0.42 }, /* Ring (LL/LL128/Simple)*/ { 0.08, 1.00, 1.00 }, /* CollNetDirect (Simple)*/ { 0.00, 0.00, 1.00 }, /* CollNetChain (Simple)*/ { 0.00, 0.00, 1.00 } },
},
.treeCorrectionFactor = {
@@ -161,18 +161,18 @@ static struct tuningModel tuning_model_2 {
static struct tuningModel tuning_model_3 {
.hwLat = {
/* NVLINK */
{ /* Tree (LL/LL128/Simple)*/ { 1.5, 1.5, 4.5 }, /* Ring (LL/LL128/Simple)*/ { 1.5, 1.5, 4.5 }, /* CollNet (LL/LL128/Simple)*/ { 1.5, 1.5, 4.5 } },
{ /* Tree (LL/LL128/Simple)*/ { 1.5, 1.5, 4.5 }, /* Ring (LL/LL128/Simple)*/ { 1.5, 1.5, 4.5 }, /* CollNetDirect (Simple)*/ { 0.0, 0.0, 4.5 }, /* CollNetChain (Simple)*/ { 0.0, 0.0, 4.5 } },
/* PCI */
{ /* Tree (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 }, /* Ring (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 }, /* CollNet (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 } },
{ /* Tree (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 }, /* Ring (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 }, /* CollNetDirect (Simple)*/ { 0.0, 0.0, 5.7 }, /* CollNetChain (Simple)*/ { 0.0, 0.0, 5.7 } },
/* NET */
{ /* Tree (LL/LL128/Simple)*/ { 17.4, 17.4, 40.3 }, /* Ring (LL/LL128/Simple)*/ { 4.1, 4.1, 40.6 }, /* CollNet (LL/LL128/Simple)*/ { 17.4, 17.4, 40.3 } },
{ /* Tree (LL/LL128/Simple)*/ { 17.4, 17.4, 40.3 }, /* Ring (LL/LL128/Simple)*/ { 4.1, 4.1, 40.6 }, /* CollNetDirect (Simple)*/ { 0.0, 0.0, 40.3 }, /* CollNetChain (Simple)*/ { 0.0, 0.0, 40.3 } },
},
.bwRatio = {
/* 2 nodes */
{ /* Tree (LL/LL128/Simple)*/ { 0.08, 1.00, 0.95 }, /* Ring (LL/LL128/Simple)*/ { 0.08, 1.00, 1.00 }, /* CollNet (LL/LL128/Simple)*/ { 1.00, 1.00, 1.00 } },
{ /* Tree (LL/LL128/Simple)*/ { 0.08, 1.00, 0.95 }, /* Ring (LL/LL128/Simple)*/ { 0.08, 1.00, 1.00 }, /* CollNetDirect (Simple)*/ { 0.00, 0.00, 1.00 }, /* CollNetChain (Simple)*/ { 0.00, 0.00, 1.00 } },
/* more than 2 nodes */
{ /* Tree (LL/LL128/Simple)*/ { 0.08, 1.00, 0.41 }, /* Ring (LL/LL128/Simple)*/ { 0.08, 1.00, 1.00 }, /* CollNet (LL/LL128/Simple)*/ { 1.00, 1.00, 1.00 } },
{ /* Tree (LL/LL128/Simple)*/ { 0.08, 1.00, 0.41 }, /* Ring (LL/LL128/Simple)*/ { 0.08, 1.00, 1.00 }, /* CollNetDirect (Simple)*/ { 0.00, 0.00, 1.00 }, /* CollNetChain (Simple)*/ { 0.00, 0.00, 1.00 } },
},
.treeCorrectionFactor = {
@@ -191,18 +191,18 @@ static struct tuningModel tuning_model_3 {
static struct tuningModel tuning_model_4 {
.hwLat = {
/* NVLINK */
{ /* Tree (LL/LL128/Simple)*/ { 0.8, 1.4, 2.5 }, /* Ring (LL/LL128/Simple)*/ { 0.8, 2.2, 3.6 }, /* CollNet (LL/LL128/Simple)*/ { 0.8, 1.4, 2.5 } },
{ /* Tree (LL/LL128/Simple)*/ { 0.8, 1.4, 2.5 }, /* Ring (LL/LL128/Simple)*/ { 0.8, 2.2, 3.6 }, /* CollNetDirect (Simple)*/ { 0.8, 1.4, 2.5 }, /* CollNetChain (Simple)*/ { 0.8, 1.4, 2.5 } },
/* PCI */
{ /* Tree (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 }, /* Ring (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 }, /* CollNet (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 } },
{ /* Tree (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 }, /* Ring (LL/LL128/Simple)*/ { 2.2, 2.2, 5.7 }, /* CollNetDirect (Simple)*/ { 0.0, 0.0, 5.7 }, /* CollNetChain (Simple)*/ { 0.0, 0.0, 5.7 } },
/* NET */
{ /* Tree (LL/LL128/Simple)*/ { 32.2, 34.4, 47.6 }, /* Ring (LL/LL128/Simple)*/ { 35.4, 87.8, 209.2 }, /* CollNet (LL/LL128/Simple)*/ { 32.2, 34.4, 47.6 } },
{ /* Tree (LL/LL128/Simple)*/ { 32.2, 34.4, 47.6 }, /* Ring (LL/LL128/Simple)*/ { 35.4, 87.8, 209.2 }, /* CollNetDirect (Simple)*/ { 0.0, 0.0, 47.6 }, /* CollNetChain (Simple)*/ { 0.0, 0.0, 47.6 } },
},
.bwRatio = {
/* 2 nodes */
{ /* Tree (LL/LL128/Simple)*/ { 0.16, 1.09, 1.61 }, /* Ring (LL/LL128/Simple)*/ { 0.15, 0.41, 1.00 }, /* CollNet (LL/LL128/Simple)*/ { 1.00, 1.00, 1.00 } },
{ /* Tree (LL/LL128/Simple)*/ { 0.16, 1.09, 1.61 }, /* Ring (LL/LL128/Simple)*/ { 0.15, 0.41, 1.00 }, /* CollNetDirect (Simple)*/ { 0.00, 0.00, 1.00 }, /* CollNetChain (Simple)*/ { 0.00, 0.00, 1.00 } },
/* more than 2 nodes */
{ /* Tree (LL/LL128/Simple)*/ { 0.16, 1.09, 1.08 }, /* Ring (LL/LL128/Simple)*/ { 0.15, 0.41, 1.00 }, /* CollNet (LL/LL128/Simple)*/ { 1.00, 1.00, 1.00 } },
{ /* Tree (LL/LL128/Simple)*/ { 0.16, 1.09, 1.08 }, /* Ring (LL/LL128/Simple)*/ { 0.15, 0.41, 1.00 }, /* CollNetDirect (Simple)*/ { 0.00, 0.00, 1.00 }, /* CollNetChain (Simple)*/ { 0.00, 0.00, 1.00 } },
},
.treeCorrectionFactor = {
@@ -232,24 +232,27 @@ static const double llMaxBws[2][3] = { /* Volta-N1/Intel-N2/Intel-N4) */ {39.0,
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;
int simpleDefaultThreads = (ringGraph->bwIntra*ringGraph->nChannels <= PCI_BW) ? 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*comm->WarpSize, NCCL_MAX_NTHREADS, simpleDefaultThreads);
comm->maxThreads[NCCL_ALGO_TREE][NCCL_PROTO_SIMPLE] = comm->maxThreads[NCCL_ALGO_COLLNET][NCCL_PROTO_SIMPLE] =
comm->maxThreads[NCCL_ALGO_TREE][NCCL_PROTO_SIMPLE] = comm->maxThreads[NCCL_ALGO_COLLNET_DIRECT][NCCL_PROTO_SIMPLE] =
getNthreads("NCCL_NTHREADS", ncclParamNthreads(), 4*comm->WarpSize, 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] =
comm->maxThreads[NCCL_ALGO_RING][NCCL_PROTO_LL] = comm->maxThreads[NCCL_ALGO_TREE][NCCL_PROTO_LL] = comm->maxThreads[NCCL_ALGO_COLLNET_DIRECT][NCCL_PROTO_LL] =
getNthreads("NCCL_NTHREADS", ncclParamNthreads(), 4*comm->WarpSize, NCCL_MAX_NTHREADS, NCCL_MAX_NTHREADS);
comm->maxThreads[NCCL_ALGO_RING][NCCL_PROTO_LL128] = comm->maxThreads[NCCL_ALGO_TREE][NCCL_PROTO_LL128] =
getNthreads("NCCL_LL128_NTHREADS", ncclParamLl128Nthreads(), 4*comm->WarpSize, NCCL_LL128_MAX_NTHREADS, NCCL_LL128_MAX_NTHREADS);
#else
getNthreads("NCCL_NTHREADS", ncclParamNthreads(), 2*WARP_SIZE, NCCL_SIMPLE_MAX_NTHREADS, simpleDefaultThreads);
comm->maxThreads[NCCL_ALGO_TREE][NCCL_PROTO_SIMPLE] =
getNthreads("NCCL_NTHREADS", ncclParamNthreads(), 2*WARP_SIZE, NCCL_SIMPLE_MAX_NTHREADS, NCCL_SIMPLE_MAX_NTHREADS);
comm->maxThreads[NCCL_ALGO_COLLNET][NCCL_PROTO_SIMPLE] = 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] =
comm->maxThreads[NCCL_ALGO_COLLNET_DIRECT][NCCL_PROTO_SIMPLE] =
comm->maxThreads[NCCL_ALGO_COLLNET_CHAIN][NCCL_PROTO_SIMPLE] = NCCL_SIMPLE_MAX_NTHREADS;
comm->maxThreads[NCCL_ALGO_RING][NCCL_PROTO_LL] = comm->maxThreads[NCCL_ALGO_TREE][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] =
comm->maxThreads[NCCL_ALGO_RING][NCCL_PROTO_LL128] = comm->maxThreads[NCCL_ALGO_TREE][NCCL_PROTO_LL128] =
getNthreads("NCCL_LL128_NTHREADS", ncclParamLl128Nthreads(), NCCL_LL128_MAX_NTHREADS/4, NCCL_LL128_MAX_NTHREADS, NCCL_LL128_MAX_NTHREADS);
#endif
int nNodes = comm->nNodes;
int nRanks = comm->nRanks;
@@ -267,7 +270,7 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
//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 };
struct ncclTopoGraph* graphs[NCCL_NUM_ALGORITHMS] = { treeGraph, ringGraph, collNetGraph, 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] = nNodes == 1 ? intraHw[a] : NCCL_HW_NET;
@@ -284,8 +287,9 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
if ((coll != ncclFuncAllReduce) && a != NCCL_ALGO_RING) continue;
for (int p=0; p<NCCL_NUM_PROTOCOLS; p++) {
float speed = nNodes <= 2 || a == NCCL_ALGO_COLLNET ? graphs[a]->speedIntra : graphs[a]->speedInter;
float busBw = comm->topo->baseBw != 0.0 ? comm->topo->baseBw : graphs[a]->nChannels * speed;
int collnet = (a == NCCL_ALGO_COLLNET_DIRECT || a == NCCL_ALGO_COLLNET_CHAIN) ? 1 : 0;
float bw = nNodes <= 2 || collnet ? graphs[a]->bwIntra : graphs[a]->bwInter;
float busBw = comm->topo->baseBw != 0.0 ? comm->topo->baseBw : graphs[a]->nChannels * bw;
// Various model refinements
#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
@@ -300,8 +304,16 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
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 : 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) {
// 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;
busBw /= factor;
}
if (a == NCCL_ALGO_COLLNET_CHAIN && p == NCCL_PROTO_SIMPLE) busBw *= .75;
#endif
if (a == NCCL_ALGO_COLLNET && p != NCCL_PROTO_SIMPLE) busBw = 0; // Oneshot CollNet only supports Simple
// Convert bus BW to algorithm BW
float ratio = (a != NCCL_ALGO_RING) ? .5 : (1.0 * nRanks) / nsteps;
@@ -309,7 +321,7 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
comm->latencies[coll][a][p] = baseLat[a][p];
float intraLat = rcclTuningModel[comm->topo->tuning].hwLat[intraHw[a]][a][p];
float interLat = rcclTuningModel[comm->topo->tuning].hwLat[NCCL_HW_NET][a][p];
float interLat = graphs[a]->latencyInter ? graphs[a]->latencyInter : rcclTuningModel[comm->topo->tuning].hwLat[NCCL_HW_NET][a][p];
//if (nNodes > 1 && p == NCCL_PROTO_LL) intraLat *= 1.8;
if (a == NCCL_ALGO_RING) {
float lat = rcclTuningModel[comm->topo->tuning].hwLat[hw[a]][a][p];
@@ -326,9 +338,11 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
} else if (a == NCCL_ALGO_TREE) {
comm->latencies[coll][a][p] +=
2 * ((nRanks/nNodes-1) * intraLat + log2i(nNodes) * interLat);
} else {
} else if (a == NCCL_ALGO_COLLNET_DIRECT) {
comm->latencies[coll][a][p] +=
2 * (std::min(1, (nRanks/nNodes-1)) * intraLat + (nRanks/nNodes-1) * 0.5) + interLat; // Add 0.5 arity serialization latency
} else if (a == NCCL_ALGO_COLLNET_CHAIN) {
comm->latencies[coll][a][p] += 2 * (nRanks/nNodes-1) * intraLat;
}
}
}
@@ -337,7 +351,7 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
// Protocols/Algorithms enable/disable, and user overrides.
// All are enabled except ll128 which is enabled by default only in certain cases.
int protoEnable[NCCL_NUM_PROTOCOLS] = { 1, 2, 1 };
int algoEnable[NCCL_NUM_ALGORITHMS] = { 1, 1, 1 };
int algoEnable[NCCL_NUM_ALGORITHMS] = { 1, 1, 1, 1 };
const char *protoStr = getenv("NCCL_PROTO");
if (protoStr) {
@@ -351,12 +365,18 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
}
// Disable CollNet if it is not supported
if (comm->collNetSupport == 0) {
algoEnable[NCCL_ALGO_COLLNET] = 0;
algoEnable[NCCL_ALGO_COLLNET_DIRECT] = 0;
algoEnable[NCCL_ALGO_COLLNET_CHAIN] = 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");
}
} else {
// Disable CollNet+Direct if not on an NVSwitch system
int nvsCount = 0;
NCCLCHECK(ncclTopoGetNvsCount(comm->topo, &nvsCount));
if (nvsCount == 0) algoEnable[NCCL_ALGO_COLLNET_DIRECT] = 0;
}
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++) {
@@ -412,13 +432,14 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
comm->threadThresholds[a][NCCL_PROTO_SIMPLE] = NCCL_SIMPLE_THREAD_THRESHOLD;
}
comm->threadThresholds[NCCL_ALGO_RING][NCCL_PROTO_LL] *= nRanks;
comm->threadThresholds[NCCL_ALGO_COLLNET][NCCL_PROTO_SIMPLE] = 256;
comm->threadThresholds[NCCL_ALGO_COLLNET_DIRECT][NCCL_PROTO_SIMPLE] = 256;
comm->threadThresholds[NCCL_ALGO_COLLNET_CHAIN][NCCL_PROTO_SIMPLE] = 256;
// Override defaults with user env
char* str = getenv("NCCL_THREAD_THRESHOLDS");
if (str) {
INFO(NCCL_ENV, "NCCL_THREAD_THRESHOLDS set by environment to %s", str);
ssize_t t[NCCL_NUM_ALGORITHMS][NCCL_NUM_PROTOCOLS] = {{ -2, -2, -2 }, { -2, -2, -2}};
ssize_t t[NCCL_NUM_ALGORITHMS][NCCL_NUM_PROTOCOLS] = {{ -2, -2, -2 }, { -2, -2, -2 }, { -2, -2, -2 }, { -2, -2, -2 }};
sscanf(str, "%ld %ld %ld %ld %ld %ld", t[0], t[0]+1, t[0]+2, t[1], t[1]+1, t[1]+2);
for (int a=0; a<NCCL_NUM_ALGORITHMS; a++) {
for (int p=0; p<NCCL_NUM_PROTOCOLS; p++) {
@@ -427,16 +448,15 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
}
}
INFO(NCCL_INIT, "threadThresholds %ld/%ld/%ld | %ld/%ld/%ld | %ld/%ld/%ld",
INFO(NCCL_INIT, "threadThresholds %ld/%ld/%ld | %ld/%ld/%ld | %ld | %ld",
comm->threadThresholds[NCCL_ALGO_TREE][NCCL_PROTO_LL],
comm->threadThresholds[NCCL_ALGO_TREE][NCCL_PROTO_LL128],
comm->threadThresholds[NCCL_ALGO_TREE][NCCL_PROTO_SIMPLE],
comm->threadThresholds[NCCL_ALGO_RING][NCCL_PROTO_LL],
comm->threadThresholds[NCCL_ALGO_RING][NCCL_PROTO_LL128],
comm->threadThresholds[NCCL_ALGO_RING][NCCL_PROTO_SIMPLE],
comm->threadThresholds[NCCL_ALGO_COLLNET][NCCL_PROTO_LL],
comm->threadThresholds[NCCL_ALGO_COLLNET][NCCL_PROTO_LL128],
comm->threadThresholds[NCCL_ALGO_COLLNET][NCCL_PROTO_SIMPLE]);
comm->threadThresholds[NCCL_ALGO_COLLNET_DIRECT][NCCL_PROTO_SIMPLE],
comm->threadThresholds[NCCL_ALGO_COLLNET_CHAIN][NCCL_PROTO_SIMPLE]);
return ncclSuccess;
}