diff options
Diffstat (limited to 'BS/dimes-hetsim-nmc.sh')
-rwxr-xr-x | BS/dimes-hetsim-nmc.sh | 78 |
1 files changed, 50 insertions, 28 deletions
diff --git a/BS/dimes-hetsim-nmc.sh b/BS/dimes-hetsim-nmc.sh index 2313985..195334b 100755 --- a/BS/dimes-hetsim-nmc.sh +++ b/BS/dimes-hetsim-nmc.sh @@ -1,9 +1,7 @@ #!/bin/bash -echo PrIM BS - mkdir -p log/$(hostname) baselines/cpu/log/$(hostname) -fn=log/$(hostname)/$(date +%Y%m%d) +fn=log/$(hostname)/dimes-hetsim-nmc # upstream DPU version uses 2048576 * uint64 ≈ 16 MiB (DPU max: 64 MiB) # upstream DPU version uses 2 queries @@ -15,19 +13,18 @@ num_queries_dpu=1048576 run_benchmark_nmc() { local "$@" + set -e 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 2>&1 - fi - return $? + make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} WITH_ALLOC_OVERHEAD=1 WITH_LOAD_OVERHEAD=1 WITH_FREE_OVERHEAD=1 INPUT_SIZE=${input_size} PROBLEM_SIZE=${num_queries} + bin/bs_host -w 0 -e 100 2>&1 } export -f run_benchmark_nmc run_benchmark_baseline() { local "$@" - OMP_NUM_THREADS=${nr_threads} ./bs_omp ${input_size} ${num_queries} ${ram} ${cpu} 2>&1 - return $? + set -e + OMP_NUM_THREADS=${nr_threads} ./bs_omp ${input_size} ${num_queries} ${ram} ${cpu} ${ram_local} ${cpu_memcpy} 2>&1 } export -f run_benchmark_baseline @@ -36,7 +33,7 @@ export -f run_benchmark_baseline echo "NMC single-node upstream-ref (1/4)" >&2 -parallel -j1 --eta --joblog ${fn}.1.joblog --header : \ +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 0 1 \ @@ -44,7 +41,7 @@ parallel -j1 --eta --joblog ${fn}.1.joblog --header : \ echo "NMC multi-node upstream-ref (2/4)" >&2 -parallel -j1 --eta --joblog ${fn}.2.joblog --header : \ +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_upstream} input_size=${input_size_upstream} \ ::: numa_rank -1 \ @@ -52,7 +49,7 @@ parallel -j1 --eta --joblog ${fn}.2.joblog --header : \ echo "NMC single-node DPU-ref (3/4)" >&2 -parallel -j1 --eta --joblog ${fn}.3.joblog --header : \ +parallel -j1 --eta --joblog ${fn}.3.joblog --resume --header : \ run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 numa_rank={numa_rank} \ num_queries=${num_queries_dpu} input_size=${input_size_dpu} \ ::: numa_rank 0 1 \ @@ -60,24 +57,51 @@ parallel -j1 --eta --joblog ${fn}.3.joblog --header : \ echo "NMC multi-node DPU-ref (4/4)" >&2 -parallel -j1 --eta --joblog ${fn}.4.joblog --header : \ +parallel -j1 --eta --joblog ${fn}.4.joblog --resume --header : \ run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 numa_rank={numa_rank} \ num_queries=${num_queries_dpu} input_size=${input_size_dpu} \ ::: numa_rank -1 \ ::: nr_dpus 1536 2048 2304 -) > ${fn}.txt - -xz -f -v -9 -M 800M ${fn}.txt +) >> ${fn}.txt cd baselines/cpu -make -B NUMA=1 ( -echo "CPU single-node upstream-ref (1/4)" >&2 +make -B NUMA=1 NUMA_MEMCPY=1 -parallel -j1 --eta --joblog ${fn}.1.joblog --header : \ +echo "CPU single-node upstream-ref with memcpy, copy node == input node (1/6)" >&2 + +parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \ + run_benchmark_baseline i={i} nr_threads={nr_threads} ram={ram} cpu={cpu} \ + ram_local={ram_local} cpu_memcpy={cpu_memcpy} \ + num_queries=${num_queries_upstream} input_size=${input_size_upstream} \ + ::: i $(seq 1 20) \ + ::: ram 0 1 \ + :::+ cpu_memcpy 0 1 \ + ::: ram_local 0 1 \ + :::+ cpu 0 1 \ + ::: nr_threads 1 2 4 8 12 16 + +echo "CPU single-node dpu-ref with memcpy, copy node == input node (2/6)" >&2 + +parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \ + run_benchmark_baseline i={i} nr_threads={nr_threads} ram={ram} cpu={cpu} \ + ram_local={ram_local} cpu_memcpy={cpu_memcpy} \ + num_queries=${num_queries_dpu} input_size=${input_size_dpu} \ + ::: i $(seq 1 20) \ + ::: ram 0 1 \ + :::+ cpu_memcpy 0 1 \ + ::: ram_local 0 1 \ + :::+ cpu 0 1 \ + ::: nr_threads 1 2 4 8 12 16 + +make -B NUMA=1 + +echo "CPU single-node upstream-ref (3/6)" >&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} \ ::: i $(seq 1 20) \ @@ -85,9 +109,9 @@ parallel -j1 --eta --joblog ${fn}.1.joblog --header : \ ::: ram 0 1 \ ::: nr_threads 1 2 4 8 12 16 -echo "CPU single-node DPU-ref (2/4)" >&2 +echo "CPU single-node DPU-ref (4/6)" >&2 -parallel -j1 --eta --joblog ${fn}.2.joblog --header : \ +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} \ ::: i $(seq 1 20) \ @@ -95,9 +119,9 @@ parallel -j1 --eta --joblog ${fn}.2.joblog --header : \ ::: ram 0 1 \ ::: nr_threads 1 2 4 8 12 16 -echo "CPU multi-node upstream-ref (3/4)" >&2 +echo "CPU multi-node upstream-ref (5/6)" >&2 -parallel -j1 --eta --joblog ${fn}.3.joblog --header : \ +parallel -j1 --eta --joblog ${fn}.5.joblog --resume --header : \ run_benchmark_baseline i={i} nr_threads={nr_threads} ram={ram} cpu={cpu} \ num_queries=${num_queries_upstream} input_size=${input_size_upstream} \ ::: i $(seq 1 20) \ @@ -105,9 +129,9 @@ parallel -j1 --eta --joblog ${fn}.3.joblog --header : \ ::: ram 0 1 \ ::: nr_threads 24 32 -echo "CPU multi-node DPU-ref (4/4)" >&2 +echo "CPU multi-node DPU-ref (6/6)" >&2 -parallel -j1 --eta --joblog ${fn}.4.joblog --header : \ +parallel -j1 --eta --joblog ${fn}.6.joblog --resume --header : \ run_benchmark_baseline i={i} nr_threads={nr_threads} ram={ram} cpu={cpu} \ num_queries=${num_queries_dpu} input_size=${input_size_dpu} \ ::: i $(seq 1 20) \ @@ -115,6 +139,4 @@ parallel -j1 --eta --joblog ${fn}.4.joblog --header : \ ::: ram 0 1 \ ::: nr_threads 24 32 -) > ${fn}.txt - -xz -f -v -9 -M 800M ${fn}.txt +) >> ${fn}.txt |