diff options
Diffstat (limited to 'BS/dimes-hetsim-nmc.sh')
| -rwxr-xr-x | BS/dimes-hetsim-nmc.sh | 39 |
1 files changed, 35 insertions, 4 deletions
diff --git a/BS/dimes-hetsim-nmc.sh b/BS/dimes-hetsim-nmc.sh index ec27ceb..7a2b12a 100755 --- a/BS/dimes-hetsim-nmc.sh +++ b/BS/dimes-hetsim-nmc.sh @@ -13,6 +13,7 @@ num_queries_dpu=1048576 run_benchmark_nmc() { local "$@" + sudo limit_ranks_to_numa_node ${numa_rank} if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} WITH_ALLOC_OVERHEAD=1 WITH_LOAD_OVERHEAD=1 WITH_FREE_OVERHEAD=1 INPUT_SIZE=${input_size} PROBLEM_SIZE=${num_queries}; then bin/bs_host -w 0 -e 100 fi @@ -31,15 +32,37 @@ export -f run_benchmark_baseline ( +echo "NMC single-node upstream-ref (1/4)" >&2 + parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \ - run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 \ + run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 numa_rank={numa_rank} \ num_queries=${num_queries_upstream} input_size=${input_size_upstream} \ - ::: nr_dpus 64 128 256 512 768 1024 1536 2048 2304 + ::: numa_rank 0 1 \ + ::: nr_dpus 64 128 256 512 768 1024 + +echo "NMC multi-node upstream-ref (2/4)" >&2 + +parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \ + run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 numa_rank={numa_rank} \ + num_queries=${num_queries_upstream} input_size=${input_size_upstream} \ + ::: numa_rank -1 \ + ::: nr_dpus 1536 2048 2304 + +echo "NMC single-node DPU-ref (3/4)" >&2 + +parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \ + run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 numa_rank={numa_rank} \ + num_queries=${num_queries_dpu} input_size=${input_size_dpu} \ + ::: numa_rank 0 1 \ + ::: nr_dpus 64 128 256 512 768 1024 + +echo "NMC multi-node DPU-ref (4/4)" >&2 parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \ - run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 \ + run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 numa_rank={numa_rank} \ num_queries=${num_queries_dpu} input_size=${input_size_dpu} \ - ::: nr_dpus 64 128 256 512 768 1024 1536 2048 2304 + ::: numa_rank -1 \ + ::: nr_dpus 1536 2048 2304 ) > ${fn}.txt @@ -50,6 +73,8 @@ make -B NUMA=1 ( +echo "CPU single-node upstream-ref (1/4)" >&2 + parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \ run_benchmark_baseline i={i} nr_threads={nr_threads} ram={ram} cpu={cpu} \ num_queries=${num_queries_upstream} input_size=${input_size_upstream} \ @@ -58,6 +83,8 @@ parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \ ::: ram 0 1 \ ::: nr_threads 1 2 4 8 12 16 32 +echo "CPU single-node DPU-ref (2/4)" >&2 + parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \ run_benchmark_baseline i={i} nr_threads={nr_threads} ram={ram} cpu={cpu} \ num_queries=${num_queries_dpu} input_size=${input_size_dpu} \ @@ -66,6 +93,8 @@ parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \ ::: ram 0 1 \ ::: nr_threads 1 2 4 8 12 16 32 +echo "CPU multi-node upstream-ref (3/4)" >&2 + parallel -j1 --eta --joblog ${fn}.3.joblog --resume --header : \ run_benchmark_baseline i={i} nr_threads={nr_threads} ram={ram} cpu={cpu} \ num_queries=${num_queries_upstream} input_size=${input_size_upstream} \ @@ -74,6 +103,8 @@ parallel -j1 --eta --joblog ${fn}.3.joblog --resume --header : \ ::: ram 0 1 \ ::: nr_threads 48 64 +echo "CPU multi-node DPU-ref (4/4)" >&2 + parallel -j1 --eta --joblog ${fn}.4.joblog --resume --header : \ run_benchmark_baseline i={i} nr_threads={nr_threads} ram={ram} cpu={cpu} \ num_queries=${num_queries_dpu} input_size=${input_size_dpu} \ |