portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-12 20:27:06 +01:00
Code Releases Activity

Optional host buffer bounce for no CUDA aware MPI

Browse Source
This commit is contained in:
Peter Boyle
2025-01-28 15:22:46 +00:00
parent 1caf8b0f86
commit 74a4f43946
7 changed files with 84 additions and 53 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
Grid/communicator/SharedMemoryMPI.cc
View File

@ -42,6 +42,11 @@ Author: Christoph Lehner <christoph@lhnr.de>
#ifdef ACCELERATOR_AWARE_MPI
#define GRID_SYCL_LEVEL_ZERO_IPC
#define SHM_SOCKETS
#else
#undef NUMA_PLACE_HOSTBUF
#ifdef NUMA_PLACE_HOSTBUF
#include <numaif.h>
#endif
#endif
#include <syscall.h>
#endif
@ -537,7 +542,30 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
// Each MPI rank should allocate our own buffer
///////////////////////////////////////////////////////////////////////////////////////////////////////////
#ifndef ACCELERATOR_AWARE_MPI
printf("Host buffer allocate for GPU non-aware MPI\n");
HostCommBuf= malloc(bytes);
#ifdef NUMA_PLACE_HOSTBUF
int numa;
char *numa_name=(char *)getenv("MPI_BUF_NUMA");
if(numa_name) {
page_size = sysconf(_SC_PAGESIZE);
numa = atoi(numa_name);
unsigned long page_count = bytes/page_size;
std::vector<void *> pages(pcount);
std::vector<int> nodes(pcount,numa);
std::vector<int> status(pcount,-1);
for(unsigned long p=0;p<page_count;p++){
pages[p] = HostCommBuf + p*page_size;
}
int ret = move_pages(0,
page_count,
&pages[0],
&nodes[0],
&status[0],
MPOL_MF_MOVE);
printf("Host buffer move to numa domain %d : move_pages returned %d\n",numa,ret);
if (ret) perror(" move_pages failed for reason:");
#endif
#endif
ShmCommBuf = acceleratorAllocDevice(bytes);
if (ShmCommBuf == (void *)NULL ) {

28
Grid/stencil/Stencil.h
View File

@ -368,6 +368,7 @@ public:
// accelerator_barrier(); // All kernels should ALREADY be complete
// _grid->StencilBarrier(); // Everyone is here, so noone running slow and still using receive buffer
// But the HaloGather had a barrier too.
#ifdef ACCELERATOR_AWARE_MPI
for(int i=0;i<Packets.size();i++){
_grid->StencilSendToRecvFromBegin(MpiReqs,
Packets[i].send_buf,
@ -376,6 +377,23 @@ public:
Packets[i].from_rank,Packets[i].do_recv,
Packets[i].xbytes,Packets[i].rbytes,i);
}
#else
#warning "Using COPY VIA HOST BUFFERS IN STENCIL"
for(int i=0;i<Packets.size();i++){
// Introduce a host buffer with a cheap slab allocator and zero cost wipe all
Packets[i].host_send_buf = _grid->HostBufferMalloc(Packets[i].xbytes);
Packets[i].host_recv_buf = _grid->HostBufferMalloc(Packets[i].rbytes);
if ( Packets[i].do_send ) {
acceleratorCopyFromDevice(Packets[i].send_buf, Packets[i].host_send_buf,Packets[i].xbytes);
}
_grid->StencilSendToRecvFromBegin(MpiReqs,
Packets[i].host_send_buf,
Packets[i].to_rank,Packets[i].do_send,
Packets[i].host_recv_buf,
Packets[i].from_rank,Packets[i].do_recv,
Packets[i].xbytes,Packets[i].rbytes,i);
}
#endif
// Get comms started then run checksums
// Having this PRIOR to the dslash seems to make Sunspot work... (!)
for(int i=0;i<Packets.size();i++){
@ -393,8 +411,16 @@ public:
else DslashLogFull();
// acceleratorCopySynchronise();// is in the StencilSendToRecvFromComplete
// accelerator_barrier();
#ifndef ACCELERATOR_AWARE_MPI
#warning "Using COPY VIA HOST BUFFERS IN STENCIL"
for(int i=0;i<Packets.size();i++){
if ( Packets[i].do_recv ) {
acceleratorCopyToDevice(Packets[i].host_recv_buf, Packets[i].recv_buf,Packets[i].rbytes);
}
}
_grid->HostBufferFreeAll();
#endif // run any checksums
_grid->StencilBarrier();
// run any checksums
for(int i=0;i<Packets.size();i++){
if ( Packets[i].do_recv )
FlightRecorder::recvLog(Packets[i].recv_buf,Packets[i].rbytes,Packets[i].from_rank);