lux.md

Sandia Lux (H100)

ssh lux27

export PATH=${PATH}:/usr/local/cuda/bin
export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/cuda/lib64
export COMM_SCOPE_SRC=$HOME/repos/comm_scope
export COMM_SCOPE_BUILD=$COMM_SCOPE_SRC/build-lux27-h100

mkdir -p $COMM_SCOPE_BUILD
cmake \
-S $COMM_SCOPE_SRC \
-B $COMM_SCOPE_BUILD \
-DCMAKE_CXX_COMPILER=g++ \
-DCMAKE_CUDA_COMPILER=nvcc \
-DSCOPE_USE_CUDA=ON \
-DSCOPE_USE_NUMA=OFF
2>&1 | tee $COMM_SCOPE_BUILD/configure.log

nice -n20 make -j$(nproc) -C $COMM_SCOPE_BUILD \
| tee $COMM_SCOPE_BUILD/build.log

CUDA Memcpy Async latency

$COMM_SCOPE_BUILD/comm_scope \
--benchmark_filter="Comm_cudaMemcpyAsync_(PinnedToGPU|GPUToPinned)/0/0" \
--benchmark_format=csv \
--benchmark_repetitions=5
--benchmark_list_tests