Rail optimization for rings (#1140)

- Modifies the ring creation algorithm to be friendlier to rail-optimized topologies (should not affect classic fabric topologies)

[ROCm/rccl commit: 4cb62f999a]

projects/rccl/tools/scripts/topo_val.sh

@@ -21,7 +21,7 @@
 
 DIR="$(cd -P "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 
-for i in {0..81}
+for i in {0..82}
 do
 	if [[ $i -eq 50 ]] || [[ $i -eq 51 ]]
 	then

projects/rccl/tools/topo_expl/topo_expl.cpp

@@ -158,6 +158,7 @@ NodeModelDesc model_descs[] = {
   {2, "topo_8p_940vm.xml",      "2 nodes gfx940 VM"},
   {2, "topo_8p_940_16n.xml",    "2 nodes gfx940 16 NICs"},
   {2, "topo_8p1h_6.xml",        "2 nodes 8P1H Alt."},
+  {5, "topo_8p_940.xml",        "5 nodes gfx940 8P"},
 };
 
 NCCL_PARAM(MaxCTAs, "MAX_CTAS", MAXCHANNELS);

projects/rccl/tools/topo_expl/utils.cpp

@@ -823,6 +823,29 @@ ncclResult_t initTransportsRank_1(struct ncclComm* comm, struct allGatherInfo *a
   // Determine local Nvls support
   //NCCLCHECK(ncclNvlsInit(comm));
 
+  // [RCCL] Compute hostIdx (based on hostHash)
+  {
+    comm->topo->nHosts = 0;
+    for (int r = 0; r < nranks; r++) {
+      int isNewHost = 1;
+      // Check if this is the first time this hostname has been used
+      for (int i = 0; i < r && isNewHost; i++) {
+        if (comm->peerInfo[i].hostHash == comm->peerInfo[r].hostHash)
+          isNewHost = 0;
+      }
+      if (isNewHost)
+      {
+        // Check if this is the same hostname associated with this rank
+        if (comm->peerInfo[r].hostHash == comm->peerInfo[rank].hostHash)
+        {
+          comm->topo->hostIdx = comm->topo->nHosts;
+          printf("Rank %d is on host %d\n", rank, comm->topo->hostIdx);
+        }
+        comm->topo->nHosts++;
+      }
+    }
+  }
+
   // Get rings and trees
   ringGraph.id = 0;
   ringGraph.pattern = NCCL_TOPO_PATTERN_RING;