portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-04-03 18:55:56 +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
Show Stats 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);

12
configure.ac
View File

@ -240,6 +240,18 @@ case ${ac_SFW_FP16} in
esac
############### MPI BOUNCE TO HOST
AC_ARG_ENABLE([accelerator-aware-mpi],
[AS_HELP_STRING([--enable-accelerator-aware-mpi=yes|no],[run mpi transfers from device])],
[ac_ACCELERATOR_AWARE_MPI=${enable_accelerator_aware_mpi}], [ac_ACCELERATOR_AWARE_MPI=yes])
case ${ac_ACCELERATOR_AWARE_MPI} in
yes)
AC_DEFINE([ACCELERATOR_CSHIFT],[1],[ Cshift runs on device])
AC_DEFINE([ACCELERATOR_AWARE_MPI],[1],[ Stencil can use device pointers]);;
*);;
esac
############### SYCL/CUDA/HIP/none
AC_ARG_ENABLE([accelerator],
[AS_HELP_STRING([--enable-accelerator=cuda|sycl|hip|none],[enable none,cuda,sycl,hip acceleration])],

32
systems/Aurora/benchmarks/bench1.pbs
View File

@ -1,6 +1,7 @@
#!/bin/bash
#PBS -q EarlyAppAccess
##PBS -q EarlyAppAccess
#PBS -q debug
#PBS -l select=1
#PBS -l walltime=00:20:00
#PBS -A LatticeQCD_aesp_CNDA
@ -12,27 +13,24 @@ source ../sourceme.sh
cp $PBS_NODEFILE nodefile
export OMP_NUM_THREADS=4
export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1
unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE
unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE
unset MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST
export MPICH_OFI_NIC_POLICY=GPU
#export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1
#unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE
#unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE
#unset MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST
#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE=0
#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE=0
#export MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST=1
export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_BUFFER_SZ=1048576
export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_THRESHOLD=131072
export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_NUM_BUFFERS_PER_CHUNK=16
export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_MAX_NUM_BUFFERS=16
export MPICH_OFI_NIC_POLICY=GPU
#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_BUFFER_SZ=1048576
#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_THRESHOLD=131072
#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_NUM_BUFFERS_PER_CHUNK=16
#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_MAX_NUM_BUFFERS=16
CMD="mpiexec -np 12 -ppn 12 -envall \
./Benchmark_dwf_fp32 --mpi 2.1.2.3 --grid 32.32.64.48 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --debug-signals"
./gpu_tile.sh ./Benchmark_dwf_fp32 --mpi 2.1.2.3 --grid 32.32.64.96 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 8 "
#for f in 1 2 3 4 5 6 7 8
for f in 1
do
echo $CMD
$CMD | tee 1node.32.32.64.48.dwf.hbm.$f
done
$CMD

34
systems/Aurora/benchmarks/gpu_tile_compact.sh
View File

@ -1,34 +0,0 @@
#!/bin/bash
#export NUMA_MAP=(2 2 2 3 3 3 2 2 2 3 3 3 )
#export NUMA_MAP=(0 0 1 1 0 0 1 1 0 0 1 1);
#export GPU_MAP=(0.0 0.1 3.0 3.1 1.0 1.1 4.0 4.1 2.0 2.1 5.0 5.1)
export NUMA_MAP=(0 0 0 0 0 0 1 1 1 1 1 1 );
export GPU_MAP=(0.0 1.0 2.0 3.0 4.0 5.0 0.1 1.1 2.1 3.1 4.1 5.1 )
export NUMA=${NUMA_MAP[$PALS_LOCAL_RANKID]}
export gpu_id=${GPU_MAP[$PALS_LOCAL_RANKID]}
unset EnableWalkerPartition
export EnableImplicitScaling=0
export ZE_AFFINITY_MASK=$gpu_id
export ONEAPI_DEVICE_FILTER=gpu,level_zero
export SYCL_PI_LEVEL_ZERO_DEVICE_SCOPE_EVENTS=0
export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE=0:5
#export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE=0:2
export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE_FOR_D2D_COPY=1
#export SYCL_PI_LEVEL_ZERO_USM_RESIDENT=1
echo "rank $PALS_RANKID ; local rank $PALS_LOCAL_RANKID ; ZE_AFFINITY_MASK=$ZE_AFFINITY_MASK ; NUMA $NUMA "
if [ $PALS_RANKID = "0" ]
then
# numactl -m $NUMA -N $NUMA onetrace --chrome-device-timeline "$@"
# numactl -m $NUMA -N $NUMA unitrace --chrome-kernel-logging --chrome-mpi-logging --chrome-sycl-logging --demangle "$@"
numactl -m $NUMA -N $NUMA "$@"
else
numactl -m $NUMA -N $NUMA "$@"
fi

2
systems/Aurora/config-command
View File

@ -17,7 +17,7 @@ export CXXFLAGS="-O3 -fiopenmp -fsycl-unnamed-lambda -fsycl -Wno-tautological-co
--with-lime=$CLIME \
--enable-shm=nvlink \
--enable-accelerator=sycl \
--enable-accelerator-aware-mpi=yes\
--enable-accelerator-aware-mpi=no\
--enable-unified=no \
MPICXX=mpicxx \
CXX=icpx

1
systems/Aurora/sourceme.sh
View File

@ -2,6 +2,7 @@
#module load mpich/icc-all-debug-pmix-gpu/52.2
#module load mpich-config/mode/deterministic
#module load intel_compute_runtime/release/821.35
module load pti-gpu
source ~/spack/share/spack/setup-env.sh
spack load c-lime