portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-10 07:55:35 +00:00
Code Releases Activity

IBM summit optimisation. Synchronise in node is still btweeen 2 halves of AC922, so could

Browse Source 
be a little faster
This commit is contained in:
Peter Boyle 2019-11-21 15:00:46 -05:00
parent ac614cbc53
commit 98ea67b636
1 changed files with 27 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

36
Grid/communicator/SharedMemoryMPI.cc
View File

@ -162,11 +162,8 @@ static inline int divides(int a,int b)
void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmDims)
{
////////////////////////////////////////////////////////////////
// Assert power of two shm_size.
// Powers of 2,3,5 only in prime decomposition for now
////////////////////////////////////////////////////////////////
int log2size = Log2Size(WorldShmSize,MAXLOG2RANKSPERNODE);
assert(log2size != -1);
int ndimension = WorldDims.size();
ShmDims=Coordinate(ndimension,1);
@ -177,7 +174,8 @@ void GlobalSharedMemory::GetShmDims(const Coordinate &WorldDims,Coordinate &ShmD
while(AutoShmSize != WorldShmSize) {
for(int p=0;p<primes.size();p++) {
int prime=primes[p];
if ( divides(prime,WorldDims[dim]/ShmDims[dim]) ) {
if ( divides(prime,WorldDims[dim]/ShmDims[dim])
&& divides(prime,WorldShmSize/AutoShmSize) ) {
AutoShmSize*=prime;
ShmDims[dim]*=prime;
break;
@ -308,7 +306,6 @@ void GlobalSharedMemory::OptimalCommunicatorHypercube(const Coordinate &processo
}
void GlobalSharedMemory::OptimalCommunicatorSharedMemory(const Coordinate &processors,Grid_MPI_Comm & optimal_comm)
{
////////////////////////////////////////////////////////////////
// Identify subblock of ranks on node spreading across dims
// in a maximally symmetrical way
@ -435,10 +432,13 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
// e.g. DGX1, supermicro board,
//////////////////////////////////////////////////////////////////////////////////////////////////////////
// cudaDeviceGetP2PAttribute(&perfRank, cudaDevP2PAttrPerformanceRank, device1, device2);
#ifdef GRID_IBM_SUMMIT
std::cout << header << "flag IBM_SUMMIT disabled CUDA set device: ensure jsrun is used correctly" <<std::endl;
// IBM Jsrun makes cuda Device numbering screwy and not match rank
std::cout << "IBM Summit or similar - NOT setting device to WorldShmRank"<<std::endl;
#else
cudaSetDevice(WorldShmRank);
std::cout << "setting device to WorldShmRank"<<std::endl;
cudaSetDevice(WorldShmRank);
#endif
///////////////////////////////////////////////////////////////////////////////////////////////////////////
// Each MPI rank should allocate our own buffer
@ -466,7 +466,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
// If it is me, pass around the IPC access key
//////////////////////////////////////////////////
cudaIpcMemHandle_t handle;
if ( r==WorldShmRank ) {
err = cudaIpcGetMemHandle(&handle,ShmCommBuf);
if ( err != cudaSuccess) {
@ -735,6 +735,24 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
std::vector<int> ranks(size); for(int r=0;r<size;r++) ranks[r]=r;
MPI_Group_translate_ranks (FullGroup,size,&ranks[0],ShmGroup, &ShmRanks[0]);
#ifdef GRID_IBM_SUMMIT
// Hide the shared memory path between sockets
// if even number of nodes
if ( (ShmSize & 0x1)==0 ) {
int SocketSize = ShmSize/2;
int mySocket = ShmRank/SocketSize;
for(int r=0;r<size;r++){
int hisRank=ShmRanks[r];
if ( hisRank!= MPI_UNDEFINED ) {
int hisSocket=hisRank/SocketSize;
if ( hisSocket != mySocket ) {
ShmRanks[r] = MPI_UNDEFINED;
}
}
}
}
#endif
SharedMemoryTest();
}
//////////////////////////////////////////////////////////////////