diff options
| -rwxr-xr-x | Microbenchmarks/CPU-DPU/splc25-alloc.sh | 35 | ||||
| -rwxr-xr-x | Microbenchmarks/CPU-DPU/splc25-transfer.sh | 36 |
2 files changed, 71 insertions, 0 deletions
diff --git a/Microbenchmarks/CPU-DPU/splc25-alloc.sh b/Microbenchmarks/CPU-DPU/splc25-alloc.sh new file mode 100755 index 0000000..6f4f055 --- /dev/null +++ b/Microbenchmarks/CPU-DPU/splc25-alloc.sh @@ -0,0 +1,35 @@ +#!/bin/bash + +mkdir -p log/$(hostname) + +run_benchmark_nmc() { + local "$@" + set -e + sudo limit_ranks_to_numa_node ${numa_rank} + ./make-size.sh ${size} + n_nops=$((size * 256)) + if make -B NR_RANKS=${nr_ranks} NR_TASKLETS=1 BL=10 DPU_BINARY=\'\"bin/dpu_size\"\' NUMA=1; then + for l in $(seq 1 20); do + bin/host_code -c ${numa_cpu} -w 1 -e 0 -x 1 -i 65536 -N $n_nops -I $(size -A bin/dpu_size | awk '($1 == ".text") {print $2/8}') + done + fi + return $? +} + +export -f run_benchmark_nmc + +for sdk in 2023.2.0 2024.1.0 2024.2.0 2025.1.0; do + + source /opt/upmem/upmem-${sdk}-Linux-x86_64/upmem_env.sh + fn=log/$(hostname)/splc25-alloc-${sdk} + + parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \ + run_benchmark_nmc nr_ranks={nr_ranks} numa_rank={numa_rank} numa_cpu={numa_cpu} size={size} \ + ::: i $(seq 1 5) \ + ::: numa_rank -1 \ + ::: numa_cpu 0 1 \ + ::: nr_ranks $(seq 1 40) \ + ::: size $(seq 0 15) \ + >> ${fn}.txt + +done diff --git a/Microbenchmarks/CPU-DPU/splc25-transfer.sh b/Microbenchmarks/CPU-DPU/splc25-transfer.sh new file mode 100755 index 0000000..0227cab --- /dev/null +++ b/Microbenchmarks/CPU-DPU/splc25-transfer.sh @@ -0,0 +1,36 @@ +#!/bin/bash + +mkdir -p log/$(hostname) + +./make-size.sh 0 + +run_benchmark_nmc() { + local "$@" + set -e + sudo limit_ranks_to_numa_node ${numa_rank} + make -B NR_RANKS=${nr_ranks} NR_TASKLETS=1 BL=10 TRANSFER=PUSH NUMA=1 + bin/host_code -a $numa_in -b $numa_out -c $numa_cpu -w 0 -e 5 -x 0 -N 0 -I $(size -A bin/dpu_code | awk '($1 == ".text") {print $2/8}') -i ${input_size} + return $? +} + +export -f run_benchmark_nmc + +# The benchmark allocates 3 * 64 * nr_ranks * 8B * input_size (one array for input, one array for output). +# With 1048576 elements (8 MiB per DPU), this gives a maximum allocation of 60 GiB, which will fit comfortably into system memory (128 GiB). +# With 2097152 elements (16 MiB per DPU), we may encounter OoM conditions, since the UPMEM SDK also allocates some memory. + +for sdk in 2023.2.0 2024.1.0 2024.2.0 2025.1.0; do + source /opt/upmem/upmem-${sdk}-Linux-x86_64/upmem_env.sh + fn=log/$(hostname)/splc25-transfer-${sdk} + + parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \ + run_benchmark_nmc nr_ranks={nr_ranks} numa_rank={numa_rank} numa_in={numa_in} numa_out={numa_out} numa_cpu={numa_cpu} input_size={input_size} \ + ::: i $(seq 1 10) \ + ::: numa_rank -1 \ + ::: numa_in 1 \ + ::: numa_out 1 \ + ::: numa_cpu 1 \ + ::: nr_ranks $(seq 1 40) \ + ::: input_size 1 1048576 \ + >> ${fn}.txt +done |