portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-04-04 11:15:55 +01:00
Code Releases Activity

Ensure different MPI ranks use different GPUs. The mapping works on Tesseract.

Browse Source
This commit is contained in:
Peter Boyle 2019-04-28 07:32:30 +01:00
parent 204a090497
commit 170ba4e619
1 changed files with 2 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
Grid/communicator/SharedMemoryMPI.cc
View File

@ -399,6 +399,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
// e.g. DGX1, supermicro board,
//////////////////////////////////////////////////////////////////////////////////////////////////////////
// cudaDeviceGetP2PAttribute(&perfRank, cudaDevP2PAttrPerformanceRank, device1, device2);
cudaSetDevice(WorldShmRank);
///////////////////////////////////////////////////////////////////////////////////////////////////////////
// Each MPI rank should allocate our own buffer
///////////////////////////////////////////////////////////////////////////////////////////////////////////
@ -411,6 +412,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
std::cerr << " SharedMemoryMPI.cc cudaMallocManaged failed NULL pointer for " << bytes<<" bytes " << std::endl;
exit(EXIT_FAILURE);
}
std::cout << " SharedMemoryMPI.cc cudaMalloc "<< bytes << "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
SharedMemoryZero(ShmCommBuf,bytes);
///////////////////////////////////////////////////////////////////////////////////////////////////////////