microsoft/mscclpp

[Perf] Two-node allreduce perf improvement

Binyang2014 opened this issue · 1 comments

Target:

n_ctx size custom (us)
1 24.0KiB 15.6
2 48.0KiB 16.6
3 72.0KiB 21.7
4 96.0KiB 17.8
8 192.0KiB 19.3
16 384.0KiB 22.7
24 576.0KiB 31.4
32 768.0KiB 28.3
48 1.1MiB 33.4
64 1.5MiB 38.4
128 3.0MiB 52.3
512 12.0MiB 139.7
1024 24.0MiB 252.7
2048 48.0MiB 476.8
4096 96.0MiB 930.2
8192 192.0MiB 1841.5

Temp update

Size (fp16) Time (us) AlgBW (GB/s) Correctness NCCL Time (us) NCCL AlgBW (GB/s) NCCL Correctness Speed Up
6.0 KiB 18.87 0.33 PASS 49.16 0.12 PASS 2.60
12.0 KiB 18.15 0.68 PASS 47.97 0.26 PASS 2.64
24.0 KiB 20.36 1.21 PASS 55.81 0.44 PASS 2.74
48.0 KiB 21.10 2.33 PASS 59.42 0.83 PASS 2.82
96.0 KiB 23.34 4.21 PASS 92.96 1.06 PASS 3.98
192.0 KiB 26.47 7.43 PASS 74.67 2.63 PASS 2.82
384.0 KiB 34.14 11.52 PASS 88.80 4.43 PASS 2.60
768.0 KiB 46.04 17.08 PASS 97.53 8.06 PASS 2.12
1.5 MiB 71.91 21.87 PASS 121.84 12.91 PASS 1.69
3.0 MiB 126.60 24.85 PASS 168.87 18.63 PASS 1.33
6.0 MiB 169.38 37.14 PASS 204.70 30.74 PASS 1.21
12.0 MiB 215.83 58.30 PASS 310.11 40.58 PASS 1.44
24.0 MiB 326.40 77.10 PASS 516.83 48.69 PASS 1.58
48.0 MiB 522.58 96.31 PASS 762.47 66.01 PASS 1.46
96.0 MiB 927.87 108.49 PASS 1324.39 76.01 PASS 1.43
192.0 MiB 1728.29 116.49 PASS 2312.49 87.06 PASS 1.34
384.0 MiB 3350.12 120.19 PASS 4352.33 92.51 PASS 1.30
768.0 MiB 6637.77 121.32 PASS 8448.14 95.32 PASS 1.27
1.5 GiB 12953.08 124.34 PASS 16360.09 98.45 PASS 1.26
3.0 GiB 25692.61 125.38 PASS 32297.60 99.74 PASS 1.26