piscina 库调研，用于 thread 预热

发表于 分类于 阅读次数：

piscina 库是一个 Node.js Worker Thread Pool，进行了一些调研。

一些特性

  • piscina 会优先使用当前负载最小的实例，所以如果有 10 个零负载的 thread，那么一下调用 10 次 runTask 可以保证每个 thread 都接收到一次。

  • piscina 支持 thread 的异步初始化，使用方式：

    1
    2
    3
    4
    5
    6
    async function initialize() {
      await someAsyncInitializationActivity();
      return ({ a, b }) => a + b;
    }

    module.exports = initialize();

  • 执行 new Piscina() 时默认不会创建 thread，会按需在调用 runTask 时创建 thread。但可以通过设置 minThreads 来使得 new Piscina() 时就创建一定数量的 thread。

运行中 thread 创建

调用 runTask 时，遇到以下两种情况，会尝试创建新的 thread：

  1. 等待队列中有正在等待的任务（表示所有 thread 都在忙碌），并且当前 thread 数量没有超过 maxThreads。
  2. 没找到可用的 thread，并且当前 thread 数量没有超过 maxThreads。

创建过程中的 thread 是不可用的，会 piscina 会等待这个 thread 异步初始化，初始化完成之后才进行使用：

1
2
3
4
5
6
7
8
9
10
11
12
worker.on('message', (message : ReadyMessage) => {
  if (message.ready === true) {
    if (workerInfo.currentUsage() === 0) {
      workerInfo.unref();
    }

    if (!workerInfo.isReady()) {
      workerInfo.markAsReady();
    }
    return;
  }
});

调度方式

寻找当前负载最小的实例：

1
2
3
4
5
6
7
8
9
10
11
12
13
findAvailable () : T | null {
  let minUsage = this.maximumUsage;
  let candidate = null;
  for (const item of this.readyItems) {
    const usage = item.currentUsage();
    if (usage === 0) return item;
    if (usage < minUsage) {
      candidate = item;
      minUsage = usage;
    }
  }
  return candidate;
}

预热 demo

我写了一个预热 piscina 的 demo：https://github.com/meixg/piscina-warmup-demo。

预热前：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
threadId: 1 done: 0th at 20.103199999779463
threadId: 2 done: 0th at 26.904699999839067
threadId: 3 done: 0th at 26.526700001209974
threadId: 5 done: 0th at 29.538399998098612
threadId: 4 done: 0th at 28.85249999910593
threadId: 6 done: 0th at 28.910900000482798
threadId: 3 done: 100th at 0.0568000003695488
threadId: 7 done: 0th at 13.932300001382828
threadId: 8 done: 0th at 22.18859999999404
threadId: 9 done: 0th at 10.147300001233816
threadId: 10 done: 0th at 15.620099999010563
threadId: 1 done: 100th at 0.08020000159740448
threadId: 2 done: 100th at 0.10360000282526016
threadId: 5 done: 100th at 0.05690000206232071
threadId: 4 done: 100th at 0.05640000104904175
threadId: 6 done: 100th at 0.06940000131726265
threadId: 7 done: 100th at 0.10370000079274178
threadId: 8 done: 100th at 0.0575999990105629
threadId: 9 done: 100th at 0.056199997663497925
threadId: 10 done: 100th at 0.07080000266432762

预热后：

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
threadId: 1 done: 0th at 1.8744000010192394
threadId: 2 done: 0th at 2.410700000822544
threadId: 3 done: 0th at 2.3161999993026257
threadId: 4 done: 0th at 2.9968000017106533
threadId: 5 done: 0th at 2.564599998295307
threadId: 6 done: 0th at 3.0161999985575676
threadId: 7 done: 0th at 3.555900000035763
threadId: 8 done: 0th at 3.5732999965548515
threadId: 9 done: 0th at 3.471400000154972
threadId: 10 done: 0th at 3.1979999989271164
threadId: 3 done: 100th at 0.057099997997283936
threadId: 5 done: 100th at 0.05829999968409538
threadId: 6 done: 100th at 0.11400000005960464
threadId: 7 done: 100th at 0.05640000104904175
threadId: 8 done: 100th at 0.08650000020861626
threadId: 9 done: 100th at 0.06109999865293503
threadId: 10 done: 100th at 0.058400001376867294
threadId: 1 done: 100th at 0.06509999930858612
threadId: 2 done: 100th at 0.055799998342990875
threadId: 4 done: 100th at 0.06909999996423721