AI/rocm-systems
2
0
Fork 0
Κώδικας Ζητήματα Pull Requests Δράσεις Πακέτα Έργα Κυκλοφορίες Wiki Δραστηριότητα

2.12.7-1

Περιήγηση Στο Πηγαίο 
Add network communication through another GPU connected with NVLink
(PXN).
Add aggregation of messages coming from different local GPUs through
PXN and going to the same destination.
Add new v5 plugin API with grouped receives and tags.
Add compat for v4 plugins.
Add naming of NCCL threads to help debugging.
Fix NVLink detection and avoid data corruption when some NVLinks are
down.
Add support for Relaxed Ordering for IB.
Add profiling and timing infrastructure.
Αυτή η υποβολή περιλαμβάνεται σε:
Sylvain Jeaugey
2022-01-07 06:39:55 -08:00
γονέας 0144073673
υποβολή 3c223c105a
70 αρχεία άλλαξαν με 6453 προσθήκες και 3430 διαγραφές
Λήψη Αρχείου Patch Λήψη Αρχείου Diff Ανάπτυξη όλων των αρχείων Σύμπτυξη όλων των αρχείων
src/graph/search.cc
+147 -74
Προβολή Αρχείου
@@ -1,5 +1,5 @@
/*************************************************************************
* Copyright (c) 2016-2021, NVIDIA CORPORATION. All rights reserved.
* Copyright (c) 2016-2022, NVIDIA CORPORATION. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
@@ -254,10 +254,10 @@ ncclResult_t ncclTopoSearchNextGpuSort(struct ncclTopoSystem* system, struct ncc
ncclResult_t ncclTopoSearchRec(struct ncclTopoSystem* system, struct ncclTopoGraph* graph, struct ncclTopoGraph* saveGraph, int* time);
// Try to keep all searchs within one second
#define NCCL_SEARCH_GLOBAL_TIMEOUT (3ULL<<19)
#define NCCL_SEARCH_TIMEOUT (1<<18)
#define NCCL_SEARCH_TIMEOUT_TREE (1<<17)
#define NCCL_SEARCH_TIMEOUT_SAMECHANNELS (1<<10)
#define NCCL_SEARCH_GLOBAL_TIMEOUT (1ULL<<18)
#define NCCL_SEARCH_TIMEOUT (1<<14)
#define NCCL_SEARCH_TIMEOUT_TREE (1<<14)
#define NCCL_SEARCH_TIMEOUT_SAMECHANNELS (1<<8)
#define FORCED_ORDER_PCI 1
#define FORCED_ORDER_REPLAY 2
@@ -305,6 +305,57 @@ ncclResult_t ncclTopoCompareGraphs(struct ncclTopoGraph* graph, struct ncclTopoG
return ncclSuccess;
}
// Build a list of the best NETs to try.
//
// "gpu" can be set to -1 to build a list suitable for all GPUs (search start) or to a given gpu
// index when trying to get back to the NIC.
//
// The list is built the following way:
// 1. Select NETs starting with those close to GPU(s), based on paths[n].type.
// 2. For each GPU, once that list of NICs with a given distance is prepared, shuffle the list
// based on the GPU NVML index so that e.g. GPU 1 chooses NIC 1 first instead of NIC 0 which
// might have been choosen by GPU 0 (case with multiple independent communicators per node)
// 3. Then add the NETs to the final list if they were not already added by another closer GPU.
ncclResult_t ncclTopoSelectNets(struct ncclTopoSystem* system, int typeInter, int gpu, int* nets, int* netCountRet) {
int netCount = 0;
int localNetCount;
int* localNets;
NCCLCHECK(ncclCalloc(&localNets, system->nodes[NET].count));
for (int t=0; t <= typeInter; t++) {
for (int g=0; g<system->nodes[GPU].count; g++) {
if (gpu != -1 && gpu != g) continue;
localNetCount = 0;
struct ncclTopoNode* gpu = system->nodes[GPU].nodes+g;
struct ncclTopoLinkList* paths = gpu->paths[NET];
for (int n=0; n<system->nodes[NET].count; n++) {
if (paths[n].type == t) localNets[localNetCount++] = n;
}
if (localNetCount == 0) continue;
// Shuffle by gpu NVML device number so that GPUs on the same PCI switch
// with multiple NICs don't use the same one as first choice.
for (int r=0; r<system->nodes[GPU].nodes[g].gpu.dev % localNetCount; r++) {
int net0 = localNets[0];
for (int i=0; i<localNetCount-1; i++) localNets[i] = localNets[i+1];
localNets[localNetCount-1] = net0;
}
// Append NICs to list
for (int i=0; i<localNetCount; i++) {
int n = localNets[i];
int found = 0;
while (nets[found] != n && found<netCount) found++;
if (found == netCount) nets[netCount++] = n;
}
}
}
*netCountRet = netCount;
free(localNets);
return ncclSuccess;
}
ncclResult_t ncclTopoSearchRecGpu(struct ncclTopoSystem* system, struct ncclTopoGraph* graph, struct ncclTopoGraph* saveGraph, struct ncclTopoNode* gpu, int step, int backToNet, int backToFirstRank, int forcedOrder, int *time) {
if ((*time) <= 0) return ncclSuccess;
(*time)--;
@@ -333,7 +384,12 @@ ncclResult_t ncclTopoSearchRecGpu(struct ncclTopoSystem* system, struct ncclTopo
int startNetIndex;
NCCLCHECK(getNetIndex(system, graph->inter[graph->nChannels*2], &startNetIndex));
struct ncclTopoNode* startNet = system->nodes[NET].nodes+startNetIndex;
for (int n=0; n<system->nodes[NET].count; n++) {
int netcount;
int* nets;
NCCLCHECK(ncclCalloc(&nets, system->nodes[NET].count));
NCCLCHECK(ncclTopoSelectNets(system, graph->typeInter, g, nets, &netcount));
for (int i=0; i<netcount; i++) {
int n = nets[i];
struct ncclTopoNode* net = system->nodes[NET].nodes+n;
if (graph->pattern == NCCL_TOPO_PATTERN_TREE && net->id != startNet->id) continue; // Trees are symmetric
if (graph->crossNic != 1 && (net->net.asic != startNet->net.asic || net->net.port != startNet->net.port)) continue;
@@ -359,6 +415,7 @@ ncclResult_t ncclTopoSearchRecGpu(struct ncclTopoSystem* system, struct ncclTopo
graph->speedInter = speedInterSave;
}
}
free(nets);
}
} else if (step < system->nodes[GPU].count-1) {
// Go to next GPU
@@ -393,65 +450,12 @@ ncclResult_t ncclTopoSearchRecGpu(struct ncclTopoSystem* system, struct ncclTopo
return ncclSuccess;
}
// Select only NICs with the maximum bandwidth w.r.t. GPUs, and sort them by distance.
ncclResult_t ncclTopoSelectNets(struct ncclTopoSystem* system, int* nets, int* netcountRet) {
float* maxwidths;
int* minhops;
int netcount = 0;
NCCLCHECK(ncclCalloc(&minhops, system->nodes[NET].count));
NCCLCHECK(ncclCalloc(&maxwidths, system->nodes[NET].count));
for (int n=0; n<system->nodes[NET].count; n++) {
maxwidths[n] = 0.0;
minhops[n] = 255;
struct ncclTopoNode* net = system->nodes[NET].nodes+n;
struct ncclTopoLinkList* paths = net->paths[GPU];
for (int g=0; g<system->nodes[GPU].count; g++) {
if (paths[g].width > maxwidths[n] || (paths[g].width == maxwidths[n] && paths[g].count < minhops[n])) {
maxwidths[n] = paths[g].width;
minhops[n] = paths[g].count;
}
}
if (netcount && maxwidths[nets[0]] > maxwidths[n]) continue; // Do not keep NICs with lower BW
if (netcount && maxwidths[nets[0]] < maxwidths[n]) netcount = 0; // Remove all NICs with lower BW
int index;
for (index = 0; index < netcount; index++) {
if (minhops[n] < minhops[nets[index]]) break;
}
// Insert net at index
// Shift all nets with higher nhops
for (int i = netcount; i>index; i--) nets[i] = nets[i-1];
// Insert this net at index
nets[index] = n;
netcount++;
}
*netcountRet = netcount;
// Then shuffle NICs with the same nhops based on the GPU device number, so that when we have
// 2 NICs and 2 GPUs and create communicators with only one GPU, we will use both NICs.
for (int start = 0; start < netcount;) {
int end = start+1;
while (end < netcount && minhops[nets[end]] == minhops[nets[start]]) end++;
// Shuffle
for (int r=0; r<system->nodes[GPU].nodes[0].gpu.dev % (end-start); r++) {
int netStart = nets[start];
for (int i=start; i<end-1; i++) nets[i] = nets[i+1];
nets[end-1] = netStart;
}
start = end;
}
free(minhops);
free(maxwidths);
return ncclSuccess;
}
ncclResult_t ncclTopoSearchRecNet(struct ncclTopoSystem* system, struct ncclTopoGraph* graph, struct ncclTopoGraph* saveGraph, int backToNet, int backToFirstRank, int* time) {
const int speed = graph->speedInter;
int* nets;
NCCLCHECK(ncclCalloc(&nets, system->nodes[NET].count));
int netcount;
NCCLCHECK(ncclTopoSelectNets(system, nets, &netcount));
NCCLCHECK(ncclTopoSelectNets(system, graph->typeInter, -1, nets, &netcount));
for (int i=0; i<netcount; i++) {
int n = nets[i];
struct ncclTopoNode* net = system->nodes[NET].nodes+n;
@@ -461,6 +465,8 @@ ncclResult_t ncclTopoSearchRecNet(struct ncclTopoSystem* system, struct ncclTopo
if (net->net.maxChannels == 0) continue;
graph->inter[graph->nChannels*2] = net->id;
graph->latencyInter = net->net.latency;
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)) {
@@ -587,7 +593,18 @@ ncclResult_t ncclTopoSearchRec(struct ncclTopoSystem* system, struct ncclTopoGra
/* User defined graph from XML file */
/************************************/
struct kvDict kvDictLinkType[] = { { "SYS", PATH_SYS }, { "PHB", PATH_PHB }, { "PIX", PATH_PIX }, { "PXB", PATH_PXB }, { "NVL", PATH_NVL }, { "NVB", PATH_NVB}, { "LOC", PATH_LOC }, { NULL, 0 } };
struct kvDict kvDictLinkType[] = {
{ "LOC", PATH_LOC },
{ "NVL", PATH_NVL },
{ "NVB", PATH_NVB },
{ "PIX", PATH_PIX },
{ "PXB", PATH_PXB },
{ "PXN", PATH_PXN },
{ "PHB", PATH_PHB },
{ "SYS", PATH_SYS },
{ NULL, 0 }
};
ncclResult_t ncclTopoGetChannelFromXml(struct ncclXmlNode *xmlChannel, int c, struct ncclTopoSystem* system, struct ncclTopoGraph* graph) {
int ngpus = system->nodes[GPU].count;
int* inter = graph->inter+2*c;
@@ -627,6 +644,7 @@ ncclResult_t ncclTopoGetGraphFromXmlSub(struct ncclXmlNode *xmlGraph, struct ncc
NCCLCHECK(xmlGetAttrInt(xmlGraph, "nchannels", &graph->nChannels));
NCCLCHECK(xmlGetAttrFloat(xmlGraph, "speedintra", &graph->speedIntra));
NCCLCHECK(xmlGetAttrFloat(xmlGraph, "speedinter", &graph->speedInter));
if (xmlGetAttrFloat(xmlGraph, "latencyinter", &graph->latencyInter) != ncclSuccess) graph->latencyInter = 0.0;
const char* str;
NCCLCHECK(xmlGetAttr(xmlGraph, "typeintra", &str));
NCCLCHECK(kvConvertToInt(str, &graph->typeIntra, kvDictLinkType));
@@ -685,6 +703,7 @@ ncclResult_t ncclTopoGetXmlFromGraph(struct ncclTopoGraph* graph, struct ncclTop
NCCLCHECK(xmlSetAttrInt(xmlGraph, "nchannels", graph->nChannels));
NCCLCHECK(xmlSetAttrFloat(xmlGraph, "speedintra", graph->speedIntra));
NCCLCHECK(xmlSetAttrFloat(xmlGraph, "speedinter", graph->speedInter));
NCCLCHECK(xmlSetAttrFloat(xmlGraph, "latencyinter", graph->latencyInter));
const char* str;
NCCLCHECK(kvConvertToStr(graph->typeIntra, &str, kvDictLinkType));
NCCLCHECK(xmlSetAttr(xmlGraph, "typeintra", str));
@@ -712,10 +731,14 @@ float speedArrayInter[] = { 48.0, 30.0, 24.0, 22.0, 18.0, 15.0, 12.0, 10.0, 9.0,
#define NSPEEDSINTRA (sizeof(speedArrayIntra)/sizeof(float))
#define NSPEEDSINTER (sizeof(speedArrayInter)/sizeof(float))
NCCL_PARAM(CrossNic, "CROSS_NIC", 2);
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->latencyInter = 0;
if (graph->crossNic == 2) graph->crossNic = 0;
graph->typeIntra = ngpus == 1 ? PATH_LOC : PATH_NVL;
graph->typeInter = PATH_PIX;
@@ -802,19 +825,13 @@ search:
goto search;
}
tmpGraph.typeIntra = ngpus == 1 ? PATH_LOC : PATH_NVL;
if (system->nodes[NET].count > 0 && tmpGraph.typeInter < PATH_SYS && (graph->nChannels == 0 || tmpGraph.typeInter < graph->typeInter || tmpGraph.typeInter < PATH_PXB)) {
if (system->nodes[NET].count > 0 && tmpGraph.typeInter < PATH_SYS && (graph->nChannels == 0 || tmpGraph.typeInter < graph->typeInter || tmpGraph.typeInter < PATH_PXN)) {
tmpGraph.typeInter += 1;
goto search;
}
tmpGraph.typeInter = PATH_PIX;
// Try a simpler tree
if (tmpGraph.pattern == NCCL_TOPO_PATTERN_SPLIT_TREE) {
tmpGraph.pattern = NCCL_TOPO_PATTERN_TREE;
goto search;
}
tmpGraph.pattern = graph->pattern;
if (crossNic && tmpGraph.crossNic == 0) {
// Try again with crossNic if permitted
tmpGraph.crossNic = crossNic;
@@ -822,6 +839,13 @@ search:
}
tmpGraph.crossNic = 0;
// Try a simpler tree
if (tmpGraph.pattern == NCCL_TOPO_PATTERN_SPLIT_TREE) {
tmpGraph.pattern = NCCL_TOPO_PATTERN_TREE;
goto 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];
@@ -915,17 +939,66 @@ ncclResult_t ncclTopoDumpGraphs(struct ncclTopoSystem* system, int ngraphs, stru
return ncclSuccess;
}
ncclResult_t ncclTopoGetNetDev(struct ncclTopoSystem* system, int rank, struct ncclTopoGraph* graph, int channelId, int rr, int* dev) {
// 0: don't use PXN for P2P, 1: use PXN if needed, 2: use PXN as much as possible to maximize aggregation
NCCL_PARAM(P2pPxnLevel, "P2P_PXN_LEVEL", 2);
#include "comm.h"
ncclResult_t ncclTopoGetNetDev(struct ncclComm* comm, int rank, struct ncclTopoGraph* graph, int channelId, int peerRank, int* dev, int* proxyRank) {
if (graph) {
// Honor the net device in the graph
int channel = channelId%graph->nChannels;
int ngpus = system->nodes[GPU].count;
int ngpus = comm->topo->nodes[GPU].count;
int index = graph->intra[channel*ngpus] == rank ? 0 : 1;
*dev = graph->inter[channel*2+index];
NCCLCHECK(ncclTopoGetIntermediateRank(comm->topo, rank, *dev, proxyRank));
} else if (peerRank == -1) {
return ncclInternalError;
} else {
int64_t id;
NCCLCHECK(ncclTopoGetLocalNet(system, rank, &id, rr));
*dev = id;
// Start with our local NIC and local Rank
NCCLCHECK(ncclTopoGetLocalNet(comm->topo, rank, dev));
*proxyRank = rank;
int pxnLevel = ncclPxnDisable() == 1 ? 0 : ncclParamP2pPxnLevel();
// See whether we can use the remote rank preferred device.
if (ncclParamCrossNic() == 0 || (pxnLevel != 0)) {
int netDev = comm->peerInfo[peerRank].netDev;
int n;
// Check that device exists on our node
if (ncclParamCrossNic() == 0) {
if (ncclTopoIdToIndex(comm->topo, NET, netDev, &n) != ncclSuccess) {
WARN("Rank %d requires NIC %d but that NIC is not available for rank %d", peerRank, netDev, rank);
return ncclInvalidUsage;
}
*dev = netDev;
}
if (pxnLevel == 1) {
int g, n;
NCCLCHECK(ncclTopoRankToIndex(comm->topo, rank, &g));
NCCLCHECK(ncclTopoIdToIndex(comm->topo, NET, netDev, &n));
struct ncclTopoNode* gpu = comm->topo->nodes[GPU].nodes+g;
if (gpu->paths[NET][n].type <= PATH_PXN) {
*dev = netDev;
NCCLCHECK(ncclTopoGetIntermediateRank(comm->topo, rank, *dev, proxyRank));
}
} else if (pxnLevel == 2) {
// Check whether we can access it through our node-local GPU for that NIC.
for (int r=0; r<comm->localRanks; r++) {
int peerRank = comm->localRankToRank[r];
if (comm->peerInfo[peerRank].netDev == netDev) {
int g1, g2, n;
NCCLCHECK(ncclTopoRankToIndex(comm->topo, rank, &g1));
NCCLCHECK(ncclTopoRankToIndex(comm->topo, peerRank, &g2));
NCCLCHECK(ncclTopoIdToIndex(comm->topo, NET, netDev, &n));
struct ncclTopoNode* peerGpu = comm->topo->nodes[GPU].nodes+g2;
if (peerGpu->paths[GPU][g1].type <= PATH_NVL && peerGpu->paths[NET][n].type <= PATH_PXB) {
*proxyRank = peerRank;
*dev = netDev;
return ncclSuccess;
}
}
}
}
}
}
return ncclSuccess;
}