blob: a14ededa8f7c97d0798ba3e3e28acfbc19a95b21 (
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
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
|
#!/bin/bash
echo PrIM BS
mkdir -p log/$(hostname) baselines/cpu/log/$(hostname)
fn=log/$(hostname)/$(date +%Y%m%d)
# upstream DPU version uses 2048576 * uint64 ≈ 16 MiB (DPU max: 64 MiB)
# upstream DPU version uses 2 queries
input_size_upstream=2048576
num_queries_upstream=2
# here: 32 MiB and 1048576 queries
input_size_dpu=$(perl -E 'say 2 ** 22')
num_queries_dpu=1048576
run_benchmark_nmc() {
local "$@"
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 $?
}
export -f run_benchmark_nmc
run_benchmark_baseline() {
local "$@"
OMP_NUM_THREADS=${nr_threads} ./bs_omp ${input_size} ${num_queries} ${ram} ${cpu} 2>&1
return $?
}
export -f run_benchmark_baseline
(
echo "NMC single-node upstream-ref (1/4)" >&2
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 \
::: nr_dpus 64 128 256 512 768 1024
echo "NMC multi-node upstream-ref (2/4)" >&2
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 \
::: nr_dpus 1536 2048 2304
echo "NMC single-node DPU-ref (3/4)" >&2
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 \
::: nr_dpus 64 128 256 512 768 1024
echo "NMC multi-node DPU-ref (4/4)" >&2
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
cd baselines/cpu
make -B NUMA=1
(
echo "CPU single-node upstream-ref (1/4)" >&2
parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \
run_benchmark_baseline i={i} nr_threads={nr_threads} ram={ram} cpu={cpu} \
num_queries=${num_queries_upstream} input_size=${input_size_upstream} \
::: i $(seq 1 20) \
::: cpu 0 1 \
::: ram 0 1 \
::: nr_threads 1 2 4 8 12 16 32
echo "CPU single-node DPU-ref (2/4)" >&2
parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \
run_benchmark_baseline i={i} nr_threads={nr_threads} ram={ram} cpu={cpu} \
num_queries=${num_queries_dpu} input_size=${input_size_dpu} \
::: i $(seq 1 20) \
::: cpu 0 1 \
::: ram 0 1 \
::: nr_threads 1 2 4 8 12 16 32
echo "CPU multi-node upstream-ref (3/4)" >&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) \
::: cpu -1 \
::: ram 0 1 \
::: nr_threads 48 64
echo "CPU multi-node DPU-ref (4/4)" >&2
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) \
::: cpu -1 \
::: ram 0 1 \
::: nr_threads 48 64
) > ${fn}.txt
xz -f -v -9 -M 800M ${fn}.txt
|
