Last Modified: a3b537c on 2026-09-12

经过一些micro opt，优

经过一些micro opt，优化后的dtact最新数据：

Spawn Tasks - Dtact:

1k mean:0.12301ms CV:26.7%

10k mean:1.6299ms CV:24.3%

100k mean:9.0277ms CV:7.8%

1M mean:69.884ms CV:7.2%

y = 0.15 * x^0.90

r = 0.997

Spawn Tasks - Tokio:

1k mean:0.54454ms CV:22.6%

10k mean:4.2720ms CV:17.1%

100k mean:32.667ms CV:28.9%

1M mean:418.93ms CV:15.7%

y = 0.49 * x^0.95

r = 0.998

Load Balance (Hot Core) - Dtact:

1k mean:0.14568ms CV:31.4%

10k mean:1.8898ms CV:18.0%

100k mean:10.260ms CV:9.2%

1M mean:99.746ms CV:6.9%

10M mean:940.77ms CV:7.5%

y = 0.16 * x^0.93

r = 0.999

Load Balance (Hot Core) - Tokio:

1k mean:0.53687ms CV:23.5%

10k mean:4.3044ms CV:16.2%

100k mean:33.673ms CV:18.2%

1M mean:460.71ms CV:18.6%

10M mean:4729.4ms CV:14.1%

y = 0.46 * x^0.99

r = 0.999

比较遗憾的是还是临时在github runner跑的，有空做一下完整benchmark。

DTA算法还在继续benchmark中，有空还得更纯粹的和work stealing对比对比跑benchmark，要不老被认为存在较大非算法因素。

History