SDCC benchmarking scripts for A100 nodes and IceLake nodes (AVX512)

benchmarks/Benchmark_dwf_fp32.cc

@@ -185,6 +185,7 @@ void Benchmark(int Ls, Coordinate Dirichlet)
   GaugeField Umu(UGrid);
   GaugeField UmuCopy(UGrid);
   SU<Nc>::HotConfiguration(RNG4,Umu);
+  //  SU<Nc>::ColdConfiguration(Umu);
   UmuCopy=Umu;
   std::cout << GridLogMessage << "Random gauge initialised " << std::endl;
 
@@ -307,6 +308,14 @@ void Benchmark(int Ls, Coordinate Dirichlet)
     if(( n2e>1.0e-4) ) {
       std::cout<<GridLogMessage << "WRONG RESULT" << std::endl;
       FGrid->Barrier();
+      std::cout<<GridLogMessage << "RESULT" << std::endl;
+      //      std::cout << result<<std::endl;
+      std::cout << norm2(result)<<std::endl;
+      std::cout<<GridLogMessage << "REF" << std::endl;
+      std::cout << norm2(ref)<<std::endl;
+      std::cout<<GridLogMessage << "ERR" << std::endl;
+      std::cout << norm2(err)<<std::endl;
+      FGrid->Barrier();
       exit(-1);
     }
     assert (n2e< 1.0e-4 );

systems/SDCC-A100/bench.slurm

@@ -0,0 +1,42 @@
+#!/bin/bash
+#SBATCH --partition csi
+#SBATCH --time=00:10:00
+#SBATCH -A csigeneral
+#SBATCH --exclusive
+#SBATCH --nodes=1
+#SBATCH --ntasks=4
+#SBATCH --qos csi
+#SBATCH --gres=gpu:4
+
+source sourceme.sh
+
+cat << EOF > select_gpu
+#!/bin/bash
+export GPU_MAP=(0 1 2 3)
+export GPU=\${GPU_MAP[\$SLURM_LOCALID]}
+export CUDA_VISIBLE_DEVICES=\$GPU
+unset ROCR_VISIBLE_DEVICES
+echo RANK \$SLURM_LOCALID using GPU \$GPU    
+exec \$*
+EOF
+chmod +x ./select_gpu
+
+
+export OMP_NUM_THREADS=4
+export OMPI_MCA_btl=^uct,openib
+export UCX_TLS=cuda,gdr_copy,rc,rc_x,sm,cuda_copy,cuda_ipc
+export UCX_RNDV_SCHEME=put_zcopy
+export UCX_RNDV_THRESH=16384
+export UCX_IB_GPU_DIRECT_RDMA=no
+export UCX_MEMTYPE_CACHE=n
+
+export OMP_NUM_THREAD=8
+#srun -N1 -n1 nvidia-smi
+#srun -N1 -n1 numactl -H > numa.txt
+srun -N1 -n1 lstopo A100-topo.pdf
+
+# 4.35 TF/s
+#srun -N1 -n1 ./benchmarks/Benchmark_dwf_fp32 --mpi 1.1.1.1 --grid 16.32.32.32 --shm 2048 --shm-mpi 0  --accelerator-threads 16
+
+srun -N1 -n4 ./select_gpu ./benchmarks/Benchmark_dwf_fp32 --mpi 1.1.2.2 --grid 32.32.64.64 --shm 2048 --shm-mpi 0  --accelerator-threads 16
+

systems/SDCC-A100/config-command

@@ -5,7 +5,7 @@
 --enable-accelerator=cuda \
 --enable-gen-simd-width=64 \
 --enable-simd=GPU \
---enable-accelerator-cshift \
+--disable-accelerator-cshift \
 --disable-fermion-reps \
 --disable-gparity \
 CXX=nvcc \

systems/SDCC-ICE/bench.slurm

@@ -0,0 +1,31 @@
+#!/bin/bash
+#SBATCH --partition lqcd
+#SBATCH --time=00:20:00
+#SBATCH -A lqcdtest
+#SBATCH --exclusive
+#SBATCH --nodes=1
+#SBATCH --ntasks=2
+#SBATCH --qos lqcd
+
+source sourceme.sh
+
+export OMP_NUM_THREAD=24
+#srun -N1 -n1 numactl -H > numa.txt
+#srun -N1 -n1 lstopo ice-topo.pdf
+
+cat << EOF > select_socket
+#!/bin/bash
+export NUM_MAP=(0 1)
+export NUMA=\${NUMA_MAP[\$SLURM_LOCALID]}
+exec \$*
+EOF
+chmod +x ./select_socket
+
+#for vol in 8.8.8.16 8.8.8.32 8.8.8.64
+#for vol in 8.8.16.16 8.8.16.32 8.8.16.64
+for vol in 8.16.16.16 8.16.16.32 8.16.16.64 16.16.16.32 16.16.16.64 24.24.24.64 32.32.32.32
+do
+srun --cpu-bind=ldoms -N1 -n2 ./select_socket ./benchmarks/Benchmark_dwf_fp32 --mpi 1.1.1.2 --grid $vol --dslash-asm > $vol.2socket.out
+srun --cpu-bind=ldoms -N1 -n1 ./select_socket ./benchmarks/Benchmark_dwf_fp32 --mpi 1.1.1.1 --grid $vol --dslash-asm > $vol.1socket.out
+done
+

systems/SDCC-ICE/config-command

@@ -1,13 +1,18 @@
 ../../configure \
---enable-comms=mpi \
+--enable-debug \
+--enable-comms=mpi-auto \
 --enable-unified=yes \
 --enable-shm=shmopen \
+--enable-shm-fast-path=shmopen \
 --enable-accelerator=none \
---enable-simd=AVX2 \
+--enable-simd=AVX512 \
 --disable-accelerator-cshift \
 --disable-fermion-reps \
 --disable-gparity \
-CXX=mpicxx \
+CXX=clang++ \
+MPICXX=mpicxx \
+LDFLAGS=-L/direct/sdcc+u/paboyle/spack/opt/spack/linux-almalinux8-icelake/gcc-8.5.0/hwloc-2.9.1-hgkscnt5pferhtde4ahctlupb6qf3vtl/lib/ \
+LIBS=-lhwloc \
 CXXFLAGS="-std=c++17"
 
 

systems/SDCC-ICE/sourceme.sh

@@ -1 +1,2 @@
+export LD_LIBRARY_PATH=/direct/sdcc+u/paboyle/spack/opt/spack/linux-almalinux8-icelake/gcc-8.5.0/llvm-12.0.1-agey6vtuw3e375rewhhobvkznjh5ltz4/lib/:$LD_LIBRARY_PATH
 module load openmpi