diff options
| -rw-r--r-- | .gitignore | 2 | ||||
| -rwxr-xr-x | VA/dimes-hetsim-hbm.sh | 10 | ||||
| -rwxr-xr-x | VA/dimes-hetsim-nmc.sh | 16 |
3 files changed, 15 insertions, 13 deletions
@@ -19,3 +19,5 @@ TRNS/baselines/cpu/trns TS/baselines/cpu/streamp_openmp TS/baselines/cpu/inputs/randomlistDPU.txt TS/baselines/cpu/inputs/randomlist10M.txt + +VA/baselines/cpu/va diff --git a/VA/dimes-hetsim-hbm.sh b/VA/dimes-hetsim-hbm.sh index acb27b1..66456e5 100755 --- a/VA/dimes-hetsim-hbm.sh +++ b/VA/dimes-hetsim-hbm.sh @@ -4,7 +4,7 @@ cd baselines/cpu make -B NUMA=1 mkdir -p log/$(hostname) -fn=log/$(hostname)/$(date +%Y%m%d)-baseline +fn=log/$(hostname)/$(date +%Y%m%d) # upstream uses 167772160 * 2 * int32 == 2.5 GiB input and 1.25 GiB output for DPU version @@ -12,8 +12,8 @@ fn=log/$(hostname)/$(date +%Y%m%d)-baseline echo "single-node execution (1/2)" >&2 -parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \ - ./va -i {input_size} -a {ram} -b {ram} -c {cpu} -t {nr_threads} -w 0 -e 20 \ +parallel -j1 --eta --joblog ${fn}.1.joblog --header : \ + ./va -i {input_size} -a {ram} -b {ram} -c {cpu} -t {nr_threads} -w 0 -e 5 \ ::: nr_threads 1 2 4 8 12 16 \ ::: cpu $(seq 0 7) \ ::: ram $(seq 0 15) \ @@ -21,8 +21,8 @@ parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \ echo "multi-node execution (2/2)" >&2 -parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \ - ./va -i {input_size} -a {ram} -b {ram} -c {cpu} -t {nr_threads} -w 0 -e 20 \ +parallel -j1 --eta --joblog ${fn}.2.joblog --header : \ + ./va -i {input_size} -a {ram} -b {ram} -c {cpu} -t {nr_threads} -w 0 -e 40 \ ::: nr_threads 32 48 64 96 128 \ ::: cpu -1 \ ::: ram $(seq 0 15) \ diff --git a/VA/dimes-hetsim-nmc.sh b/VA/dimes-hetsim-nmc.sh index 3d94fba..f622b47 100755 --- a/VA/dimes-hetsim-nmc.sh +++ b/VA/dimes-hetsim-nmc.sh @@ -20,7 +20,7 @@ export -f run_benchmark_nmc echo "NMC single-node operation (1/2)" >&2 -parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \ +parallel -j1 --eta --joblog ${fn}.1.joblog --header : \ run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 input_size={input_size} numa_rank={numa_rank} \ ::: numa_rank 0 1 \ ::: nr_dpus 64 128 256 512 768 1024 \ @@ -28,9 +28,9 @@ parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \ echo "NMC multi-node operation (2/2)" >&2 -parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \ +parallel -j1 --eta --joblog ${fn}.2.joblog --header : \ run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 input_size={input_size} numa_rank={numa_rank} \ - ::: numa_rank -1 \ + ::: numa_rank any \ ::: nr_dpus 1536 2048 2304 \ ::: input_size 167772160 @@ -45,20 +45,20 @@ make -B NUMA=1 echo "CPU single-node operation (1/2)" >&2 -parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \ +parallel -j1 --eta --joblog ${fn}.1.joblog --header : \ ./va -i {input_size} -a {ram} -b {ram} -c {cpu} -t {nr_threads} -w 0 -e 40 \ ::: ram 0 1 \ ::: cpu 0 1 \ - ::: nr_threads 1 2 4 8 12 16 32 \ + ::: nr_threads 1 2 4 8 12 16 \ ::: input_size 167772160 echo "CPU multi-node operation (1/2)" >&2 -parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \ +parallel -j1 --eta --joblog ${fn}.2.joblog --header : \ ./va -i {input_size} -a {ram} -b {ram} -c {cpu} -t {nr_threads} -w 0 -e 40 \ - ::: ram any \ + ::: ram 0 1 \ ::: cpu -1 \ - ::: nr_threads 48 64 \ + ::: nr_threads 24 32 \ ::: input_size 167772160 ) > ${fn}.txt |