blob: bffba028ff5e6009a867d0d160235c3ce120cc70 (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
|
#!/bin/bash
mkdir -p log/$(hostname)
fn=log/$(hostname)/dimes-hetsim-transfer
./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.
(
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 0 1 \
::: numa_in 0 1 \
::: numa_out 0 1 \
::: numa_cpu 0 1 \
::: nr_ranks $(seq 1 20) \
::: input_size 1 1048576
parallel -j1 --eta --joblog ${fn}.2.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 any \
::: numa_in all \
::: numa_out all \
::: numa_cpu 0 1 \
::: nr_ranks $(seq 21 40) \
::: input_size 1 1048576
) >> ${fn}.txt
|
