Gracefully parallelize Node.js on multi-cores and multi-nodes


stars 🌟issues ⚠️updated 🛠created 🐣size 🏋️‍♀️
70May 21, 2017Mar 31, 2017Minified + gzip package size for accel in KB



Accelerate your Node.js codes on multicores and multi-nodes with no efforts!

Quick Start

Install accel.js: (root privilege may be required if npm is configured to store packages in /usr)

npm install -g accel

Write a program with computational intensive functions

function cpuIntensive(n) {
    var result=0;
    for (var i=0; i<n; i++) result=Math.random();
    return result;

for (var i=10000; i<20000; i++) {
    var result=cpuIntensive(i);

And accelerate it on your awesome dual core machine!

var accel=require("accel");
var $=accel();

function cpuIntensive(n) {
    var result=0;
    for (var i=0; i<n; i++) result=Math.random();
    return result;

for (var i=10000; i<20000; i++) {
    $.cpuIntensive(i, (result) => {

Broadcasting variables? Easy!

var accel=require("accel");
var $=accel();

$(function generatePseudoRandom(next) {
    next = next * $.magicNumber1 + $.magicNumber2;
    return Math.floor(next/65536) % 32768;

$.generatePseudoRandom(50, (res) => {

Asynchronous function works well, too:

var accel=require("accel");
var $=accel();

$.magicTricks=function(str) {...};
$(function readAndCalculate(callback) {
    fs.readFile('/etc/passwd', 'utf8', (err, data) => {
}, "async");

$.readAndCalculate((res) => {


Launch via acceljs

accel.js provides an easy tools for launching parallelized node.js codes, simply use

acceljs yourNodeJSCode.js

is okay. By the means, there's no need to launch accel.js inside your application. For more information, use acceljs --help to view usage.

Launch in your own app (NOT RECOMMENDED)

Sometimes your may not want to use acceljs to launch your parallelized app (why not?). Then you can launch it in your codes, via accel.init(), which should be called exactly once, you can specify how many local workers (run on local CPUs) and how many remote workers (run on other machines) are in the workers pool.

accel.init(options, [callback])

  • options: specify the options for accel.js
    • env (optional): List, each element is one of the following: (default is {local: -1})
      • {local: workerCount}: add some number of local workers in the pool, if workerCount=-1, it is set to the number of CPUs in local machine.
      • [Not supported yet] {remote: workerCount, endPoint: "someEndpoint"}: add some number of remote workers on the endpoint in the pool.
  • callback: called when all workers are spawned, and connections are established.


  • Do all the operations in callback to ensure everything is done

Broadcast Scope

Broadcast scope is the center of all magic. In your code which is executed remotely, all the local variables are not accessible (of course!) and you can use broadcast scope to make them available.

var $=accel(); will create a new broadcast scope. It is recommended to use separate scope for every .js file to achieve scope isolations. When functions in some scope are executed remotely, the scope will be synchronized automatically and incrementally to broadcast the variables.

The broadcast scope is nothing more than a js-object, just read/write values by $.someKey=someValue or $["someKey"]=someValue.

The scope can also be used as a function, in order to register worker-functions:

Register worker-functions: $(someFunction, [mode])

  • someFunction: a function that will be executed remotely. If the function has name, it can be called later by accessing the broadcast scope ($.funcName(...)); if the function is anonymous, it can only be called at once ($(function(){xxx})(...)).
  • mode (optional): define the mode of the function, default is "sync"
    • "sync": the function is working in a synchronous way, its result will be returned as soon as the function exit.
    • "async": the function is an asynchronous one, and abide the convention that the last parameter is the callback. The result will be passed by the parameter of callback.


  • Synchronizing broadcast scope is not free, so reusing variables helps keeping programs fast. Anonymous function is a one-time function and will never be reused, therefore not recommended.

  • Syncing is achieved by JSON, therefore unable to reserve prototype chain, function-in-object, etc., although function directly as value will be reserved (as shown below). To use nested objects with prototype and functions remotely please use homogeneous dependency.

    $.x1={func: function() {...}};        // func will not be reserved
    $.x2=function() {...};                  // good to go!

Invoke worker-functions

For synchronous function, pass all the parameters as normal, plus a trailing function as callback:

function someFunction(arg1, arg2) {}
$.someFunction(arg1, arg2, (res) => {
    // now the function has completed

For asynchronous function, call the function as it is.

function someAsyncFunc(arg1, arg2, callback) {}
$(someAsyncFunc, "async");
$.someAsyncFunc(arg1, arg2, (params) => {
    // now the function has completed

Besides, all the remote functions can be decorated before called:

$.someFunction(arg1, arg2, (res) => { });    // normal invocation
$.someFunction.urgent()(arg1, arg2, (res) => { });    // urgent invocation, the task will be scheduled first
$, arg2, (res) => { });    // targeted invocation, the task will be assigned to node 2
$.someFunction.toAll()(arg1, arg2, (res) => { });    // cohert invocation, each node will execute this task exactly once

Homogeneous Dependency

In most cases, it is required to import some packages outside in worker functions. This is achieved by introducing homogeneous dependency:


$.os.platform();        // you can invoke it locally
$(function() {
    return $.os.platform();    // or invoke it remotely
})((res) => {
    // do something

Note: Only use it after assigning to a broadcast scope.

$.os=accel.require("os"); is OK, while var os=accel.require("os"); is not ok.


Accel.js also has wonderful support for MPI jobs. After launched, the workers talk to each other by channel.

Broadcast Channel

Broadcast Channel is a Golang style multi-consumer-multi-producer channel that is accessible by all the workers as well as master. Everyone can act as a producer and a consumer, i.e., send to the channel or receive from the channel. The receiving function will not callback before anything is received, while the sending function will not callback when the buffer is full.

Use the simple API to create a broadcast channel:

$.c=new accel.Channel([bufferSize]);

Where the parameter is the size of channel, which by default is 0.

Unicast Channel

Unicast Channel has the same API as broadcast channel (see below), but only support the communication between two which are specified in creation. Since unicast channel utilize direct IPC between workers instead of being controlled by master, it has lower latency.

Use similar API to create a unicast channel:

$.c=new accel.Channel(participant1, participant2[, bufferSize]);

In which participant is the number of workers, or -1 for master.

Channel API

After creation (and potentially synchronizing via broadcast scope), channel can be accessed by Recv and Send (for broadcast, everyone can access the channel, while for unicast channel, non-specified access results in exception)

Channel.Send(obj[, callback])

  • obj: a js-object that will be serialized and sent.
  • callback (optinal): called when the object is successfully received or enters the buffer of channel.


  • callback (optional): called when a new object is received, which is passed as the first parameter. If no callback is presented, it only tries to receive an object and discards it.


Repeatedly receiving messages, calling callback and deciding whether to receive next based on the return value of callback.

  • callback: called each time when a new message is received. And if callback returns true, channel will start to wait for the next message; terminate waiting otherwise.

If you find any bugs or have a feature request, please open an issue on github!

The npm package download data comes from npm's download counts api and package details come from