diff options
Diffstat (limited to 'Microbenchmarks')
| -rw-r--r-- | Microbenchmarks/CPU-DPU/host/app.c | 4 | ||||
| -rwxr-xr-x | Microbenchmarks/CPU-DPU/run-transfer-rank.sh | 2 |
2 files changed, 3 insertions, 3 deletions
diff --git a/Microbenchmarks/CPU-DPU/host/app.c b/Microbenchmarks/CPU-DPU/host/app.c index ba63d89..504d0de 100644 --- a/Microbenchmarks/CPU-DPU/host/app.c +++ b/Microbenchmarks/CPU-DPU/host/app.c @@ -170,9 +170,9 @@ int main(int argc, char **argv) { stop(&timer, 3); if (rep >= p.n_warmup) { - printf("[::] transfer UPMEM | n_dpus=%d n_ranks=%d n_tasklets=%d n_nops=%d n_instr=%d e_type=%s n_elements=%u e_mode=%s" + printf("[::] transfer UPMEM | n_dpus=%d n_ranks=%d n_tasklets=%d n_nops=%d n_instr=%d e_type=%s n_elements=%u n_elements_per_dpu=%u e_mode=%s" " | latency_dram_mram_ns=%lu latency_mram_dram_ns=%lu throughput_dram_mram_Bps=%f throughput_mram_dram_Bps=%f", - nr_of_dpus, nr_of_ranks, NR_TASKLETS, p.n_nops, p.n_instr, XSTR(T), transfer_size, transfer_mode, + nr_of_dpus, nr_of_ranks, NR_TASKLETS, p.n_nops, p.n_instr, XSTR(T), transfer_size, transfer_size / NR_DPUS, transfer_mode, timer.nanoseconds[1], timer.nanoseconds[3], transfer_size * sizeof(T) * 1e9 / timer.nanoseconds[1], transfer_size * sizeof(T) * 1e9 / timer.nanoseconds[3]); diff --git a/Microbenchmarks/CPU-DPU/run-transfer-rank.sh b/Microbenchmarks/CPU-DPU/run-transfer-rank.sh index 169c5a0..9bb58bc 100755 --- a/Microbenchmarks/CPU-DPU/run-transfer-rank.sh +++ b/Microbenchmarks/CPU-DPU/run-transfer-rank.sh @@ -14,7 +14,7 @@ for i in 1 4 8 16 32 48 64; do # 8 B ... 64 MB for l in 4096 65536 262144 1048576 4194304 6291456 8388608; do make -B NR_DPUS=$i NR_TASKLETS=1 BL=10 TRANSFER=$k - bin/host_code -w 0 -e 100 -x 1 -i $l + bin/host_code -w 0 -e 100 -x 1 -N 0 -I $(size -A bin/dpu_code | awk '($1 == ".text") {print $2/8}') -i $l done done for k in SERIAL PUSH; do |