portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-09-19 16:55:37 +01:00
Code Releases Activity

Options to bounce through a host buffer if

Browse Source 
--disable-accelerator-aware-mpi
This commit is contained in:
Peter Boyle 2024-03-26 00:37:19 +00:00
parent 434c3e7f1d
commit 1f53458af8
3 changed files with 44 additions and 43 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

21
Grid/communicator/SharedMemory.cc
View File

@ -40,6 +40,7 @@ int GlobalSharedMemory::_ShmAlloc;
uint64_t GlobalSharedMemory::_ShmAllocBytes;
std::vector<void *> GlobalSharedMemory::WorldShmCommBufs;
void * GlobalSharedMemory::HostCommBuf;
Grid_MPI_Comm GlobalSharedMemory::WorldShmComm;
int GlobalSharedMemory::WorldShmRank;
@ -66,6 +67,26 @@ void GlobalSharedMemory::SharedMemoryFree(void)
/////////////////////////////////
// Alloc, free shmem region
/////////////////////////////////
#ifndef ACCELERATOR_AWARE_MPI
void *SharedMemory::HostBufferMalloc(size_t bytes){
void *ptr = (void *)host_heap_top;
host_heap_top += bytes;
host_heap_bytes+= bytes;
if (host_heap_bytes >= host_heap_size) {
std::cout<< " HostBufferMalloc exceeded heap size -- try increasing with --shm <MB> flag" <<std::endl;
std::cout<< " Parameter specified in units of MB (megabytes) " <<std::endl;
std::cout<< " Current alloc is " << (bytes/(1024*1024)) <<"MB"<<std::endl;
std::cout<< " Current bytes is " << (host_heap_bytes/(1024*1024)) <<"MB"<<std::endl;
std::cout<< " Current heap is " << (host_heap_size/(1024*1024)) <<"MB"<<std::endl;
assert(host_heap_bytes<host_heap_size);
}
return ptr;
}
void SharedMemory::HostBufferFreeAll(void) {
host_heap_top =(size_t)HostCommBuf;
host_heap_bytes=0;
}
#endif
void *SharedMemory::ShmBufferMalloc(size_t bytes){
// bytes = (bytes+sizeof(vRealD))&(~(sizeof(vRealD)-1));// align up bytes
void *ptr = (void *)heap_top;

16
Grid/communicator/SharedMemory.h
View File

@ -75,7 +75,9 @@ public:
static int Hugepages;
static std::vector<void *> WorldShmCommBufs;
#ifndef ACCELERATOR_AWARE_MPI
static void *HostCommBuf;
#endif
static Grid_MPI_Comm WorldComm;
static int WorldRank;
static int WorldSize;
@ -120,6 +122,13 @@ private:
size_t heap_bytes;
size_t heap_size;
#ifndef ACCELERATOR_AWARE_MPI
size_t host_heap_top; // set in free all
size_t host_heap_bytes;// set in free all
void *HostCommBuf; // set in SetCommunicator
size_t host_heap_size; // set in SetCommunicator
#endif
protected:
Grid_MPI_Comm ShmComm; // for barriers
@ -151,7 +160,10 @@ public:
void *ShmBufferTranslate(int rank,void * local_p);
void *ShmBufferMalloc(size_t bytes);
void ShmBufferFreeAll(void) ;
#ifndef ACCELERATOR_AWARE_MPI
void *HostBufferMalloc(size_t bytes);
void HostBufferFreeAll(void);
#endif
//////////////////////////////////////////////////////////////////////////
// Make info on Nodes & ranks and Shared memory available
//////////////////////////////////////////////////////////////////////////

50
Grid/communicator/SharedMemoryMPI.cc
View File

@ -512,46 +512,6 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
// Hugetlbfs mapping intended
////////////////////////////////////////////////////////////////////////////////////////////
#if defined(GRID_CUDA) ||defined(GRID_HIP) || defined(GRID_SYCL)
//if defined(GRID_SYCL)
#if 0
void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
{
void * ShmCommBuf ;
assert(_ShmSetup==1);
assert(_ShmAlloc==0);
//////////////////////////////////////////////////////////////////////////////////////////////////////////
// allocate the pointer array for shared windows for our group
//////////////////////////////////////////////////////////////////////////////////////////////////////////
MPI_Barrier(WorldShmComm);
WorldShmCommBufs.resize(WorldShmSize);
///////////////////////////////////////////////////////////////////////////////////////////////////////////
// Each MPI rank should allocate our own buffer
///////////////////////////////////////////////////////////////////////////////////////////////////////////
ShmCommBuf = acceleratorAllocDevice(bytes);
if (ShmCommBuf == (void *)NULL ) {
std::cerr << " SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
exit(EXIT_FAILURE);
}
std::cout << WorldRank << Mheader " SharedMemoryMPI.cc acceleratorAllocDevice "<< bytes
<< "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
SharedMemoryZero(ShmCommBuf,bytes);
assert(WorldShmSize == 1);
for(int r=0;r<WorldShmSize;r++){
WorldShmCommBufs[r] = ShmCommBuf;
}
_ShmAllocBytes=bytes;
_ShmAlloc=1;
}
#endif
#if defined(GRID_CUDA) ||defined(GRID_HIP) ||defined(GRID_SYCL)
void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
{
void * ShmCommBuf ;
@ -574,6 +534,9 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
///////////////////////////////////////////////////////////////////////////////////////////////////////////
// Each MPI rank should allocate our own buffer
///////////////////////////////////////////////////////////////////////////////////////////////////////////
#ifndef ACCELERATOR_AWARE_MPI
HostCommBuf= malloc(bytes);
#endif
ShmCommBuf = acceleratorAllocDevice(bytes);
if (ShmCommBuf == (void *)NULL ) {
std::cerr << " SharedMemoryMPI.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
@ -738,7 +701,6 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
_ShmAllocBytes=bytes;
_ShmAlloc=1;
}
#endif
#else
#ifdef GRID_MPI3_SHMMMAP
@ -962,6 +924,12 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
}
ShmBufferFreeAll();
#ifndef ACCELERATOR_AWARE_MPI
host_heap_size = heap_size;
HostCommBuf= GlobalSharedMemory::HostCommBuf;
HostBufferFreeAll();
#endif
/////////////////////////////////////////////////////////////////////
// find comm ranks in our SHM group (i.e. which ranks are on our node)
/////////////////////////////////////////////////////////////////////