benchmark v2.1.0 vs v2.0.26 2025‐09‐21
Due to the removal of the unnecessary atomic counter in the benchmark function, all scores are higher than the previous benchmarks. So we rerun benchmarks for v2.0 (v2.0.26), to know how much performance boost in v2.1.
In the charted image, "crossfire" is v2.1.0 (v2.0.26 scores has been excluded).
This time, we have added benchmark results of tokio::mpsc and async-channel.
- Intel(R) Core(TM) i7-8550U 4 cores with 8 hyper-threads, turbo boost on
- Ubuntu 24.04
- rust 1.87.0
- tokio 1.47.1
- crossbeam v0.8.4
- flume 0.11.1
- kanal 0.1.1
unit: million/s
| mpsc bounded size 100 tokio | 1x1 | 2x1 | 4x1 | 8x1 | 16x1 |
|---|---|---|---|---|---|
| crossfire::spsc (v2.1.0) | 37.116 | ||||
| crossfire::mpsc (v2.1.0) | 36.488 | 38.821 | 38.591 | 35.093 | 33.811 |
| crossfire::mpmc (v2.1.0) | 35.456 | 38.902 | 37.981 | 34.914 | 33.462 |
| crossfire::spsc (v2.0.26) | 29.332 | ||||
| crossfire::mpsc (v2.0.26) | 28.444 | 30.417 | 29.889 | 23.823 | 22.909 |
| crossfire::mpmc (v2.0.26) | 27.122 | 31.419 | 30.194 | 24.028 | 23.145 |
| kanal | 10.549 | 10.767 | 10.381 | 10.186 | 10.462 |
| flume | 7.9363 | 5.925 | 4.2317 | 2.6682 | 2.0159 |
| async-channel | 6.0784 | 5.5772 | 5.0109 | 4.2245 | 3.6917 |
| tokio | 6.4089 | 5.9596 | 4.847 | 3.4356 | 2.6008 |
| mpmc bounded size 100 tokio | 2x2 | 4x4 | 8x8 | 16x16 |
|---|---|---|---|---|
| crossfire::mpmc (v2.1.0) | 36.566 | 33.565 | 29.185 | 29.388 |
| crossfire::mpmc (v2.0.26) | 29.573 | 25.409 | 19.694 | 19.868 |
| kanal | 23.578 | 7.6766 | 17.652 | 19.327 |
| flume | 5.5305 | 3.8072 | 3.1198 | 2.8173 |
| async-channel | 5.677 | 4.9866 | 3.9711 | 3.764 |
unit: million/s
| mpsc bounded size 1 tokio | 1x1 | 2x1 | 4x1 | 8x1 | 16x1 |
|---|---|---|---|---|---|
| crossfire::spsc (v2.1) | 5.4666 | ||||
| crossfire::mpsc (v2.1) | 5.4725 | 5.3884 | 5.3196 | 5.4943 | 6.419 |
| crossfire::mpmc (v2.1) | 5.1895 | 5.4266 | 5.1654 | 5.4779 | 6.4887 |
| kanal | 0.32689 | 0.38732 | 0.55014 | 0.81729 | 0.97879 |
| flume | 0.17626 | 0.24744 | 0.42052 | 0.56692 | 0.53564 |
| async-channel | 0.108 | 0.091666 | 0.082779 | 0.08482 | 0.085164 |
(The result of v2.0.26 is roughly the same as v2.1.0)
unit: million/s
| mpsc unbounded tokio | 1x1 | 2x1 | 4x1 | 8x1 | 16x1 |
|---|---|---|---|---|---|
| crossfire::spsc (v2.1.0) | 35.26 | ||||
| crossfire::mpsc (v2.1.0) | 35.54 | 37.859 | 35.043 | 35.707 | 35.451 |
| crossfire::mpmc (v2.1.0) | 35.599 | 36.817 | 35.138 | 35.436 | 35.985 |
| crossfire::spsc (v2.0.26) | 23.803 | ||||
| crossfire::mpsc (v2.0.26) | 23.899 | 29.339 | 27.604 | 29.411 | 29.195 |
| crossfire::mpmc (v2.0.26) | 24.01 | 29.9 | 27.733 | 29.138 | 28.987 |
| kanal | 36.054 | 38.284 | 32.155 | 28.497 | 31.057 |
| flume | 12.279 | 11.153 | 9.1671 | 7.5639 | 7.4911 |
| async-channel | 5.8421 | 5.0894 | 4.9306 | 4.2715 | 4.2476 |
| tokio | 5.1505 | 5.8724 | 8.1421 | 8.1421 | 7.9821 |
unit: million/s
| mpmc ubounded tokio | 2x2 | 4x4 | 8x8 | 16x16 | |
|---|---|---|---|---|---|
| crossfire::mpmc (v2.1.0) | 36.787 | 34.391 | 34.438 | 34.026 | |
| crossfire::mpmc (v2.0.26) | 29.6 | 28.002 | 27.229 | 27.643 | |
| kanal | 29.896 | 19.624 | 16.905 | 18.758 | |
| flume | 7.4458 | 4.3728 | 4.3213 | 4.5644 | |
| async-channel | 4.9271 | 4.5423 | 4.1087 | 4.3246 |
unit: million/s
| mpsc bounded size 100 blocking | 1x1 | 2x1 | 4x1 | 8x1 | 16x1 |
|---|---|---|---|---|---|
| crossfire::spsc | 50.297 | ||||
| crossfire::mpsc | 49.308 | 50.976 | 47.487 | 43.966 | 39.538 |
| crossfire::mpmc | 48.365 | 48.336 | 49.379 | 47.541 | 39.919 |
| crossbeam | 30.81 | 31.817 | 28.523 | 33.881 | 33.278 |
| kanal | 24.679 | 30.299 | 23.54 | 18.938 | 17.71 |
| flume | 13.132 | 8.0159 | 3.3441 | 1.0544 | 0.51756 |
| mpmc bounded size 100 blocking | 2x2 | 4x4 | 8x8 | 16x16 |
|---|---|---|---|---|
| crossfire::mpmc | 49.187 | 46.557 | 28.752 | 15.906 |
| crossbeam | 33.082 | 26.475 | 19.569 | 13.245 |
| kanal | 16.104 | 20.459 | 7.9588 | 9.6356 |
| flume | 6.8004 | 3.775 | 2.5744 | 2.0051 |
unit: million/s
| mpsc bounded size 1 blocking | 1x1 | 2x1 | 4x1 | 8x1 | 16x1 |
|---|---|---|---|---|---|
| crossfire::spsc | 4.3325 | ||||
| crossfire::mpsc | 4.3841 | 4.5029 | 5.2273 | 4.2731 | 4.4604 |
| crossfire::mpmc | 4.383 | 4.426 | 5.1816 | 4.7245 | 3.8124 |
| crossbeam | 4.8877 | 4.7712 | 5.5047 | 5.5022 | 5.8586 |
| kanal | 2.7992 | 2.9495 | 2.4648 | 1.6127 | 1.4414 |
| flume | 0.4225 | 0.50063 | 0.55053 | 0.39121 | 0.30535 |
unit: million/s
| mpsc unbounded blocking | 1x1 | 2x1 | 4x1 | 8x1 | 16x1 |
|---|---|---|---|---|---|
| crossfire(spsc) | 40.057 | ||||
| crossfire(mpsc) | 40.259 | 39.197 | 38.479 | 37.964 | 37.871 |
| crossfire(mpmc) | 40.187 | 37.998 | 39.35 | 37.007 | 38.424 |
| crossbeam | 28.095 | 41.119 | 38.08 | 39.705 | 39.771 |
| kanal | 28.146 | 39.225 | 37.825 | 23.933 | 24.65 |
| flume | 12.445 | 10.086 | 9.1263 | 7.3567 | 7.2426 |
| mpmc unbounded blocking | 2x2 | 4x4 | 8x8 | 16x16 |
|---|---|---|---|---|
| crossfire(mpmc) | 39.022 | 37.306 | 38.913 | 30.729 |
| crossbeam | 40.445 | 37.37 | 39.01 | 38.949 |
| kanal | 17.218 | 21.637 | 22.024 | 14.515 |
| flume | 8.1359 | 4.8267 | 4.1219 | 4.2847 |
-
In async context, crossfire v2.1 achieves +26% performance, by removing the notification layer (which is of no use for us) in crossbeam-channel.
-
In blocking context, crossfire v2.1 bounded size-100 surpasses crossbeam, because crossfire only wakes one waker after one message ideally, while crossbeam wakes all senders/receivers, and optimizes the functions in crossbeam-queue to reduce moving in full condition. Meanwhile, for size-1 case, crossfire is slightly slower than crossbeam; although the backoff pattern is the same, the reason is still unknown.
-
kanal unbounded channel might grow the allocation aggressively, without shrinking, a bit faster than crossbeam. Crossfire is roughly the same as Kanal, although we use the same queue as Crossbeam, slightly faster due to the removal of the notification layer.
-
The speed of Kanal fluctuated during benchmark, non-regularly changing upon increasing the concurrency.