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

Pipeline mode getting better -- 2 nodes @ 10TF/s per node on Aurora

Browse Source
This commit is contained in:
Peter Boyle 2025-01-29 09:22:21 +00:00
parent 74a4f43946
commit d6b2727f86
9 changed files with 112 additions and 78 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
Grid/communicator/Communicator_base.h
View File

@ -136,7 +136,7 @@ public:
for(int d=0;d<_ndimension;d++){
column.resize(_processors[d]);
column[0] = accum;
std::vector<CommsRequest_t> list;
std::vector<MpiCommsRequest_t> list;
for(int p=1;p<_processors[d];p++){
ShiftedRanks(d,p,source,dest);
SendToRecvFromBegin(list,
@ -166,8 +166,8 @@ public:
////////////////////////////////////////////////////////////
// Face exchange, buffer swap in translational invariant way
////////////////////////////////////////////////////////////
void CommsComplete(std::vector<CommsRequest_t> &list);
void SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void CommsComplete(std::vector<MpiCommsRequest_t> &list);
void SendToRecvFromBegin(std::vector<MpiCommsRequest_t> &list,
void *xmit,
int dest,
void *recv,

79
Grid/communicator/Communicator_mpi3.cc
View File

@ -317,7 +317,7 @@ void CartesianCommunicator::GlobalSumVector(double *d,int N)
assert(ierr==0);
}
void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void CartesianCommunicator::SendToRecvFromBegin(std::vector<MpiCommsRequest_t> &list,
void *xmit,
int dest,
void *recv,
@ -342,7 +342,7 @@ void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &lis
assert(ierr==0);
list.push_back(xrq);
}
void CartesianCommunicator::CommsComplete(std::vector<CommsRequest_t> &list)
void CartesianCommunicator::CommsComplete(std::vector<MpiCommsRequest_t> &list)
{
int nreq=list.size();
@ -361,7 +361,7 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
int from,
int bytes)
{
std::vector<CommsRequest_t> reqs(0);
std::vector<MpiCommsRequest_t> reqs(0);
unsigned long xcrc = crc32(0L, Z_NULL, 0);
unsigned long rcrc = crc32(0L, Z_NULL, 0);
@ -404,6 +404,29 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
int from,int dor,
int xbytes,int rbytes,int dir)
{
/*
* Bring sequence from Stencil.h down to lower level.
* Assume using XeLink is ok
#warning "Using COPY VIA HOST BUFFERS IN STENCIL"
// 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);
}
for(int i=0;i<Packets.size();i++){
if ( Packets[i].do_recv ) {
}
}
_grid->HostBufferFreeAll();
*/
int ncomm =communicator_halo.size();
int commdir=dir%ncomm;
@ -421,28 +444,60 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
double off_node_bytes=0.0;
int tag;
void * host_recv = NULL;
void * host_xmit = NULL;
if ( dor ) {
if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
tag= dir+from*32;
#ifdef ACCELERATOR_AWARE_MPI
ierr=MPI_Irecv(recv, rbytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
assert(ierr==0);
list.push_back(rrq);
#else
host_recv = this->HostBufferMalloc(rbytes);
ierr=MPI_Irecv(host_recv, rbytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
assert(ierr==0);
CommsRequest_t srq;
srq.PacketType = InterNodeRecv;
srq.bytes = rbytes;
srq.req = rrq;
srq.host_buf = host_recv;
srq.device_buf = recv;
list.push_back(srq);
#endif
off_node_bytes+=rbytes;
}
} else{
#ifdef NVLINK_GET
void *shm = (void *) this->ShmBufferTranslate(from,xmit);
assert(shm!=NULL);
acceleratorCopyDeviceToDeviceAsynch(shm,recv,rbytes);
#endif
}
}
if (dox) {
// rcrc = crc32(rcrc,(unsigned char *)recv,bytes);
if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
tag= dir+_processor*32;
#ifdef ACCELERATOR_AWARE_MPI
ierr =MPI_Isend(xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
assert(ierr==0);
list.push_back(xrq);
#else
std::cout << " send via host bounce "<<std::endl;
host_xmit = this->HostBufferMalloc(xbytes);
acceleratorCopyFromDevice(xmit, host_xmit,xbytes);
ierr =MPI_Isend(host_xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
assert(ierr==0);
CommsRequest_t srq;
srq.PacketType = InterNodeXmit;
srq.bytes = xbytes;
srq.req = xrq;
srq.host_buf = host_xmit;
srq.device_buf = xmit;
list.push_back(srq);
#endif
off_node_bytes+=xbytes;
} else {
#ifndef NVLINK_GET
@ -463,11 +518,25 @@ void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsReque
acceleratorCopySynchronise();
if (nreq==0) return;
#ifdef ACCELERATOR_AWARE_MPI
std::vector<MPI_Status> status(nreq);
int ierr = MPI_Waitall(nreq,&list[0],&status[0]);
assert(ierr==0);
list.resize(0);
#else
// Wait individually and immediately copy receives to device
// Promition to Asynch copy and single wait is easy
MPI_Status status;
for(int r=0;r<nreq;r++){
int ierr = MPI_Wait(&list[r].req,&status);
assert(ierr==0);
if ( list[r].PacketType==InterNodeRecv ) {
acceleratorCopyToDevice(list[r].host_buf,list[r].device_buf,list[r].bytes);
}
}
list.resize(0);
this->HostBufferFreeAll();
#endif
}
void CartesianCommunicator::StencilBarrier(void)
{

14
Grid/communicator/SharedMemory.h
View File

@ -46,8 +46,22 @@ NAMESPACE_BEGIN(Grid);
#if defined (GRID_COMMS_MPI3)
typedef MPI_Comm Grid_MPI_Comm;
typedef MPI_Request MpiCommsRequest_t;
#ifdef ACCELERATOR_AWARE_MPI
typedef MPI_Request CommsRequest_t;
#else
enum PacketType_t { InterNodeXmit, InterNodeRecv, IntraNodeXmit, IntraNodeRecv };
typedef struct {
PacketType_t PacketType;
void *host_buf;
void *device_buf;
unsigned long bytes;
MpiCommsRequest_t req;
} CommsRequest_t;
#endif
#else
typedef int MpiCommsRequest_t;
typedef int CommsRequest_t;
typedef int Grid_MPI_Comm;
#endif

2
Grid/communicator/SharedMemoryMPI.cc
View File

@ -543,7 +543,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
///////////////////////////////////////////////////////////////////////////////////////////////////////////
#ifndef ACCELERATOR_AWARE_MPI
printf("Host buffer allocate for GPU non-aware MPI\n");
HostCommBuf= malloc(bytes);
HostCommBuf= malloc(bytes); /// CHANGE THIS TO malloc_host
#ifdef NUMA_PLACE_HOSTBUF
int numa;
char *numa_name=(char *)getenv("MPI_BUF_NUMA");

4
Grid/lattice/PaddedCell.h
View File

@ -467,8 +467,8 @@ public:
send_buf.resize(buffer_size*2*depth);
recv_buf.resize(buffer_size*2*depth);
std::vector<CommsRequest_t> fwd_req;
std::vector<CommsRequest_t> bwd_req;
std::vector<MpiCommsRequest_t> fwd_req;
std::vector<MpiCommsRequest_t> bwd_req;
int words = buffer_size;
int bytes = words * sizeof(vobj);

27
Grid/stencil/Stencil.h
View File

@ -368,7 +368,6 @@ 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,
@ -377,23 +376,6 @@ 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++){
@ -411,15 +393,6 @@ 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();
for(int i=0;i<Packets.size();i++){
if ( Packets[i].do_recv )

2
systems/Aurora/benchmarks/bench1.pbs
View File

@ -29,7 +29,7 @@ export MPICH_OFI_NIC_POLICY=GPU
CMD="mpiexec -np 12 -ppn 12 -envall \
./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 "
--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 8 "
echo $CMD
$CMD

52
systems/Aurora/benchmarks/bench2.pbs
View File

@ -1,58 +1,36 @@
#!/bin/bash
#PBS -q EarlyAppAccess
##PBS -q EarlyAppAccess
#PBS -q debug
#PBS -l select=2
#PBS -l walltime=00:20:00
#PBS -A LatticeQCD_aesp_CNDA
#export OMP_PROC_BIND=spread
#unset OMP_PLACES
cd $PBS_O_WORKDIR
source ../sourceme.sh
#module load pti-gpu
cp $PBS_NODEFILE nodefile
export OMP_NUM_THREADS=4
export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1
export MPICH_OFI_NIC_POLICY=GPU
#export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1
#unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE
#unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE
#unset MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST
export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE=0
export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE=0
export MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST=1
export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_BUFFER_SZ=1048576
export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_THRESHOLD=131072
export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_NUM_BUFFERS_PER_CHUNK=16
export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_MAX_NUM_BUFFERS=16
export MPICH_OFI_NIC_POLICY=GPU
# 12 ppn, 2 nodes, 24 ranks
#
CMD="mpiexec -np 24 -ppn 12 -envall \
./gpu_tile.sh \
./Benchmark_comms_host_device --mpi 2.2.2.3 --grid 24.32.32.24 \
--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 32"
#$CMD | tee 2node.comms.hbm
#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
CMD="mpiexec -np 24 -ppn 12 -envall \
./Benchmark_dwf_fp32 --mpi 2.2.2.3 --grid 32.32.64.48 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap --debug-signals"
./gpu_tile.sh ./Benchmark_dwf_fp32 --mpi 2.2.2.3 --grid 32.64.64.96 \
--shm-mpi 0 --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
$CMD | tee 2node.32.32.64.48.dwf.hbm.$f
done
CMD="mpiexec -np 24 -ppn 12 -envall \
./gpu_tile.sh \
./Benchmark_dwf_fp32 --mpi 2.2.2.3 --grid 64.64.64.96 \
--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
#$CMD | tee 2node.64.64.64.96.dwf.hbm
$CMD

4
systems/Aurora/benchmarks/gpu_tile.sh
View File

@ -28,8 +28,8 @@ echo "rank $PALS_RANKID ; local rank $PALS_LOCAL_RANKID ; ZE_AFFINITY_MASK=$ZE_A
if [ $PALS_RANKID = "0" ]
then
# numactl -m $NUMAM -N $NUMAP unitrace --chrome-kernel-logging --chrome-mpi-logging --chrome-sycl-logging --demangle "$@"
numactl -m $NUMAM -N $NUMAP "$@"
numactl -m $NUMAM -N $NUMAP unitrace --chrome-kernel-logging --chrome-mpi-logging --chrome-sycl-logging --demangle "$@"
# numactl -m $NUMAM -N $NUMAP "$@"
else
numactl -m $NUMAM -N $NUMAP "$@"
fi