speculum - transform concurrently

speculum is a stream multiplier. This Node.js package implements an interface that wraps multiple transform streams into a single, concurrently processing, stream.Transform. In use cases where result order is not paramount, speculum can reduce run time.

An IO-heavy transform stream’s run time T grows linearly with the number N of chunks (units of IO work) C:

T = N * C

speculum divides the run time by the number of concurrent streams X:

T = N * C / X

So far for theory.

Example

Here is a, somewhat contrived but runnable, example comparing the run times from single to ten concurrent streams:

'use strict'

// example - measure one to ten concurrent streams

const assert = require('assert')
const speculum = require('speculum')
const stream = require('stream')
const util = require('util')

// A transform that does asynchronous work.
util.inherits(Echo, stream.Transform)
function Echo (opts) {
  if (!(this instanceof Echo)) {
    return new Echo(opts)
  }
  stream.Transform.call(this, opts)
}
Echo.prototype._transform = function (chunk, enc, cb) {
  setTimeout(() => {
    this.push(chunk)
    cb()
  }, 100)
}

// An input stream to read from.
util.inherits(Count, stream.Readable)
function Count (opts, max) {
  stream.Readable.call(this, opts)
  this.count = 0
  this.max = max
}
Count.prototype._read = function () {
  let ok = false
  do {
    ok = this.push(String(this.count++))
  } while (this.count < this.max && ok)
  if (this.count >= this.max) {
    this.push(null)
  }
}

// Leverage x streams to transform, delayed echoing in this example, data from
// our readable stream.
function run (x, cb) {
  const s = speculum(null, Echo, x)
  s.on('end', cb)
  s.on('error', cb)

  const reader = new Count({ highWaterMark: 0 }, 10)

  reader.pipe(s).resume()
}

function measure (x, cb) {
  function time (t) {
    return t[0] * 1e9 + t[1]
  }
  const t = process.hrtime()
  run(x, (er) => {
    const lat = time(process.hrtime(t))
    console.log(x + ' X took ' + (lat / 1e6).toFixed(2) + ' ms')
    cb(er)
  })
}

(function go (max, x = 1) {
  if (x > max) return
  measure(x, (er) => {
    assert(!er, er)
    go(max, x + 1)
  })
})(10)

You can run this with:

$ node example.js

On this MacBook Air (11-inch, Mid 2011), with Node v6.7.0, I get:

1 X took 1088.49 ms
2 X took 525.67 ms
3 X took 411.52 ms
4 X took 317.99 ms
5 X took 210.97 ms
6 X took 210.99 ms
7 X took 210.68 ms
8 X took 210.24 ms
9 X took 211.61 ms
10 X took 104.25 ms

Considerations

Clearly, we have to balance workload and overhead to use this efficiently. Specifically, we need an idea of how many chunks our stream may need to process, before we can choose an effective number of concurrent streams. But efficiency, of course, varies depending on duration and consistence of the work to be done inside our multiplied stream.

speculum is a good fit if you want to reduce run time by leveraging existing transform streams concurrently, not minding unordered output. In other use cases, where you might be writing a concurrent transform stream from scratch, try throughv.

Exports

speculum(opts, create, x = 1)

• opts Object() | null | undefined Options passed to the stream constructor.
• create function A constructor function applied to create transform streams.
• x Number() | null | undefined The number of concurrent transform streams defaults to one.

The speculum module exports a function that returns an instance of the Speculum class which extends stream.Transform. To access the Speculum class require('speculum'). The speculum stream round-robins transformers, constructed by create, and exposes their buffers and errors through its stream.Readable interface.

Installation

With npm, do:

$ npm install speculum

License

MIT License