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

Fix bug #307 : wrong NIC selection on the reduction tree.

Browse Source 
The reduction tree (tree up) was inverting the NICs to use,
causing performance issue in cases where we are using different
NICs on a given channel.
This commit is contained in:
Sylvain Jeaugey
2020-04-09 17:14:07 -07:00
parent 533e3702cf
commit b5b6c6acdd
4 changed files with 10 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graph/search.cc
+5 -2
View File
@@ -804,7 +804,10 @@ ncclResult_t ncclTopoDumpGraphs(struct ncclTopoSystem* system, int ngraphs, stru
return ncclSuccess;
}
ncclResult_t ncclTopoGetNetDev(struct ncclTopoGraph* graph, int dir, int channelId, int* dev) {
*dev = graph->inter[(channelId%graph->nChannels)*2+dir];
ncclResult_t ncclTopoGetNetDev(struct ncclTopoSystem* system, struct ncclTopoGraph* graph, int rank, int channelId, int* dev) {
int channel = channelId%graph->nChannels;
int ngpus = system->nodes[GPU].count;
int index = graph->intra[channel*ngpus] == rank ? 0 : 1;
*dev = graph->inter[channel*2+index];
return ncclSuccess;
}