Q test suite nextTick results

q-nexttick-times.md

Comparing Different Implementations of process.nextTick using the Q test suite.

Used: q-spec.js@a1a416.

Implementations compared:

Based on MessageChannel

This is currently in Q:

// http://www.nonblocking.io/2011/06/windownexttick.html
var channel = new MessageChannel();
// linked list of tasks (single, with head node)
var head = {}, tail = head;
channel.port1.onmessage = function () {
    head = head.next;
    var task = head.task;
    delete head.task;
    task();
};
nextTick = function (task) {
    tail = tail.next = {task: task};
    channel.port2.postMessage(0);
};

Notably Firefox does not have a MessageChannel interface, so it falls back to setTimeout(task, 0).

Based on postMessage

There were reports of this being faster, so I tried the following implementation:

var MESSAGE = "q-nextTick-" + Math.random();

// linked list of tasks (single, with head node)
var head = {}, tail = head;

var onGlobalMessage = function (event) {
    // This will catch all incoming messages (even from other
    // windows), so make sure we're getting a message we triggered.
    if (event.source === window && event.data === MESSAGE) {
        head = head.next;
        var task = head.task;
        delete head.task;
        task();
    }
};

if (window.addEventListener) {
    window.addEventListener("message", onGlobalMessage, false);
} else {
    window.attachEvent("onmessage", onGlobalMessage);
}

nextTick = function (task) {
    tail = tail.next = {task: task};
    postMessage(MESSAGE, "*");
};

Based on msSetImmediate

In IE10, there is a msSetImmediate function available, based on the Efficient Script Yielding draft. Notably, it must be called as a method of window, so we need to do

nextTick = msSetImmediate.bind(window);

Based on setTimeout

In Firefox, MessageChannel is not available, so currently in Q it falls back to

nextTick = function (task) {
    setTimeout(task, 0);
};

Results

I ran the Q test suite three times in a variety of settings. Here are the results, in seconds. Not amazingly scientific, but still data.

Firefox postMessage:               4.61, 3.72, 4.63
Firefox setTimeout:                3.34, 3.51, 3.46

Chrome postMessage:                1.25, 1.30, 1.31
Chrome MessageChannel:             1.30, 1.30, 1.29

IE10 in IE9 mode, postMessage:     1.08, 1.07, 1.08
IE10 in IE9 mode, MessageChannel:  1.64, 1.70, 1.65
IE10 in IE10 mode, msSetImmediate: 0.99, 1.02, 1.03

The results indicate a few interesting things. Note that all these conclusions really only apply to the Q test suite, which may not necessarily be a good general test suite for efficient script yielding.

• postMessage and MessageChannel are not appreciably different in Chrome. In IE9 postMessage has a slight advantage.
• Using postMessage in Firefox is actually slower than using setTimeout (!)
• msSetImmediate is just barely faster than postMessage.

Thus, to conclude, it seems Q should stay with its current strategy, which is to only try MessageChannel before moving back to setTimeout.