portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-13 01:05:36 +00:00
Code Releases Activity

Comments on further topology discovery work

Browse Source
This commit is contained in:
Peter Boyle 2018-09-11 04:20:04 +01:00
parent 7bf82f5b37
commit 94d721a20b
1 changed files with 6 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
lib/communicator/SharedMemoryMPI.cc
View File

@ -199,6 +199,12 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
MPI_Barrier(WorldShmComm); MPI_Barrier(WorldShmComm);
WorldShmCommBufs.resize(WorldShmSize); 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 // Each MPI rank should allocate our own buffer
/////////////////////////////////////////////////////////////////////////////////////////////////////////// ///////////////////////////////////////////////////////////////////////////////////////////////////////////