From 94d721a20b177dfc14fd624d6e6f7cc000d6add8 Mon Sep 17 00:00:00 2001 From: Peter Boyle Date: Tue, 11 Sep 2018 04:20:04 +0100 Subject: [PATCH] Comments on further topology discovery work --- lib/communicator/SharedMemoryMPI.cc | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/lib/communicator/SharedMemoryMPI.cc b/lib/communicator/SharedMemoryMPI.cc index 3ef5f5c7..44939332 100644 --- a/lib/communicator/SharedMemoryMPI.cc +++ b/lib/communicator/SharedMemoryMPI.cc @@ -199,6 +199,12 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags) MPI_Barrier(WorldShmComm); WorldShmCommBufs.resize(WorldShmSize); + ////////////////////////////////////////////////////////////////////////////////////////////////////////// + // TODO/FIXME : NOT ALL NVLINK BOARDS have full Peer to peer connectivity. + // The annoyance is that they have partial peer 2 peer. This occurs on the 8 GPU blades. + // e.g. DGX1, supermicro board, + ////////////////////////////////////////////////////////////////////////////////////////////////////////// + // cudaDeviceGetP2PAttribute(&perfRank, cudaDevP2PAttrPerformanceRank, device1, device2); /////////////////////////////////////////////////////////////////////////////////////////////////////////// // Each MPI rank should allocate our own buffer ///////////////////////////////////////////////////////////////////////////////////////////////////////////