[rccl-tests] Add topo mapping file

projects/rccl-tests/X7950H0-BW1000/mpirun_rccltest

@@ -133,9 +133,9 @@ else
   if [ ${#copyto_hosts[@]} -gt 0 ]; then
     echo "[WRAPPER] Copying files to remote nodes in parallel: ${copyto_hosts[*]}"
     for node in "${copyto_hosts[@]}"; do
-      rsync -az -e "ssh -p ${ssh_port}" ${PWD}/build ${PWD}/scripts ${rccltest_runscript} ${NCCL_TOPO_FILE} ${NCCL_GRAPH_FILE} "${node}:${PWD}/" &
-      rsync -az -e "ssh -p ${ssh_port}" /opt/dtk/rccl/lib ${node}:/opt/dtk/rccl/ &
-      rsync -az -e "ssh -p ${ssh_port}" /opt/mpi /opt/ucx ${node}:/opt/ &
+      rsync -azP -e "ssh -p ${ssh_port}" ${PWD}/build ${PWD}/scripts ${rccltest_runscript} ${NCCL_TOPO_FILE} ${NCCL_GRAPH_FILE} ${NCCL_TOPO_MAPPING_FILE} "${node}:${PWD}/" &
+      rsync -azP -e "ssh -p ${ssh_port}" /opt/dtk/rccl/lib ${node}:/opt/dtk/rccl/ &
+      rsync -azP -e "ssh -p ${ssh_port}" /opt/mpi /opt/ucx ${node}:/opt/ &
     done
     wait
     echo "[WRAPPER] Files synchronized successfully."

projects/rccl-tests/X7950H0-BW1000/run-allreduce.sh

@@ -28,7 +28,8 @@ export NCCL_NET_GDR_READ=1
 # export NCCL_PROTO=Simple
 export NCCL_SIMPLE_CHANNELS=32
 unset NCCL_NCHANNELS_PER_PEER
-export NCCL_TOPO_FILE=${PWD}/topo-gdr-bw1000.xml
+export NCCL_TOPO_MAPPING_FILE=${PWD}/topo-mapping-bw1000.xml
+# export NCCL_TOPO_FILE=${PWD}/topo-gdr-bw1000.xml # topo 和 topo mapping 二选一即可
 # export NCCL_GRAPH_FILE=${PWD}/graph-16r-allreduce.xml
 ./mpirun_rccltest -np 2 \
   ./build/all_reduce_perf -b 4 -e 16G -f 2 -w 3 -g 1

projects/rccl-tests/X7950H0-BW1000/run-alltoall.sh

@@ -26,7 +26,8 @@ export NCCL_P2P_LEVEL=SYS
 export NCCL_NET_GDR_LEVEL=PHB
 export NCCL_NET_GDR_READ=1
 unset NCCL_NCHANNELS_PER_PEER
-export NCCL_TOPO_FILE=${PWD}/topo-gdr-bw1000.xml
+export NCCL_TOPO_MAPPING_FILE=${PWD}/topo-mapping-bw1000.xml
+# export NCCL_TOPO_FILE=${PWD}/topo-gdr-bw1000.xml # topo 和 topo mapping 二选一即可
 ./mpirun_rccltest -np 2 \
   ./build/alltoall_perf -b 32 -e 16G -f 2 -w 3 -g 1
 ./mpirun_rccltest -np 4 \

projects/rccl-tests/X7950H0-BW1000/run-p2p.sh

@@ -16,7 +16,8 @@ export NCCL_P2P_LEVEL=SYS
 export NCCL_NET_GDR_LEVEL=PHB
 export NCCL_NET_GDR_READ=1
 unset NCCL_NCHANNELS_PER_PEER
-export NCCL_TOPO_FILE=${PWD}/topo-gdr-bw1000.xml
+export NCCL_TOPO_MAPPING_FILE=${PWD}/topo-mapping-bw1000.xml
+# export NCCL_TOPO_FILE=${PWD}/topo-gdr-bw1000.xml # topo 和 topo mapping 二选一即可
 
 for g in {0..7}; do
 echo 

projects/rccl-tests/X7950H0-BW1000/topo-mapping-bw1000.xml

+<system version="2">
+  <group name="gfx936_8_x86_64_HygonGenuine_mlx5_11_InfiniBand_40-200-200-200-200-40-5-200-200-200-200_1_8_1">
+    <cpu numaid="3">
+      <pci>
+        <nic id="mlx5_1"/>
+        <nic id="mlx5_2"/>
+        <gpu dev="0"/>
+        <gpu dev="1"/>
+      </pci>
+    </cpu>
+    <cpu numaid="0">
+      <pci>
+        <nic id="mlx5_3"/>
+        <nic id="mlx5_4"/>
+        <gpu dev="2"/>
+        <gpu dev="3"/>
+      </pci>
+    </cpu>
+    <cpu numaid="7">
+      <pci>
+        <nic id="mlx5_7"/>
+        <nic id="mlx5_8"/>
+        <gpu dev="4"/>
+        <gpu dev="5"/>
+      </pci>
+    </cpu>
+    <cpu numaid="4">
+      <pci>
+        <nic id="mlx5_9"/>
+        <nic id="mlx5_10"/>
+        <gpu dev="6"/>
+        <gpu dev="7"/>
+      </pci>
+    </cpu>
+  </group>
+</system>