在Javascript（V8）中，为什么数组上的forEach比简单的for循环消耗更多的内存？

Question

I'm performing some simple data validation on a large set of data in Node.js (version v7.5.0, with a matrix of 15849x12771 entries). The entire data set is in memory for now, for performance reasons. Therefore it is critical for me to reduce the amount of memory consumed to a theoretical minimum (each number representing 8 bytes in JS).

Please compare the following ways of achieving the same thing.

with forEach

  regressData.forEach((yxa, yxaIndex) => {
    yxa.forEach((yx, yxIndex) => {
      if (!_.isFinite(yx)) {
        throw new Error(`non-finite entry at [${yxaIndex}, ${yxIndex}]`);
      }
    });
  });

This consumes all of my node process' memory at 4GB+, causing it to never (until my patience runs out anyway) finish the loop (I guess it will use slower swap memory).

And then the identical version with a typical for :

  for (var yxai = 0, yxal = regressData.length; yxai < yxal; yxai++) {
    const yx = regressData[yxai];
    for (var yxi = 0, yxl = yx.length; yxi < yxl; yxi++) {
      if (!_.isFinite(yx[yxi])) {
        throw new Error(`non-finite entry at [${yxai}, ${yxi}]`);
      }
    }
  }

This consumes virtually no extra memory, causing the validation to be done in less than a second.

Is this behavior as expected? I had anticipated that because the forEach s have closed scopes there would be no issues of additional memory usage when compared to a traditional for loop.

EDIT: standalone test

node --expose-gc test_foreach.js

if (!gc) throw new Error('please run node like node --expose-gc test_foreach.js');

const _ = require('lodash');

// prepare data to work with

const x = 15849;
const y = 12771;

let regressData = new Array(x);
for (var i = 0; i < x; i++) {
  regressData[i] = new Array(y);
  for (var j = 0; j < y; j++) {
    regressData[i][j] = _.random(true);
  }
}

// for loop
gc();
const mb_pre_for = _.round(process.memoryUsage().heapUsed / 1024 / 1024, 2);
console.log(`memory consumption before for loop ${mb_pre_for} megabyte`);
validateFor(regressData);
gc();
const mb_post_for = _.round(process.memoryUsage().heapUsed / 1024 / 1024, 2);
const mb_for = _.round(mb_post_for - mb_pre_for, 2);
console.log(`memory consumption by for loop ${mb_for} megabyte`);

// for each loop
gc();
const mb_pre_foreach = _.round(process.memoryUsage().heapUsed / 1024 / 1024, 2);
console.log(`memory consumption before foreach loop ${mb_pre_foreach} megabyte`);
validateForEach(regressData);
gc();
const mb_post_foreach = _.round(process.memoryUsage().heapUsed / 1024 / 1024, 2);
const mb_foreach = _.round(mb_post_foreach - mb_pre_foreach, 2);
console.log(`memory consumption by foreach loop ${mb_foreach} megabyte`);

function validateFor(regressData) {
  for (var yxai = 0, yxal = regressData.length; yxai < yxal; yxai++) {
    const yx = regressData[yxai];
    for (var yxi = 0, yxl = yx.length; yxi < yxl; yxi++) {
      if (!_.isFinite(yx[yxi])) {
        throw new Error(`non-finite entry at [${yxai}, ${yxi}]`);
      }
    }
  }
};

function validateForEach(regressData) {
  regressData.forEach((yxa, yxaIndex) => {
    yxa.forEach((yx, yxIndex) => {
      if (!_.isFinite(yx)) {
        throw new Error(`non-finite entry at [${yxaIndex}, ${yxIndex}]`);
      }
    });
  });
};

Output:

toms-mbp-2:mem_test tommedema$ node --expose-gc test_foreach.js
memory consumption before for loop 1549.31 megabyte
memory consumption by for loop 0.31 megabyte
memory consumption before foreach loop 1549.66 megabyte
memory consumption by foreach loop 3087.9 megabyte

Answer 1

(V8 developer here.) This is an unfortunate consequence of how Array.forEach is implemented in V8's old execution pipeline (full codegen + Crankshaft). In short, what happens is that under some circumstances, using forEach on an array changes the internal representation of that array to a much less memory efficient format. (Specifically: if the array contained only double values before, and forEach has also been used on arrays with elements of other types but not too many different kinds of objects, and the code runs hot enough to get optimized. It's fairly complicated ;-) )

With the new execution pipeline (currently behind the --future flag, will be turned on by default soon), I'm no longer seeing this additional memory consumption.

(That said, classic for loops do tend to have a small performance advantage over forEach , just because there's less going on under the hood (per ES spec). In many real workloads, the difference is too small to matter, but in microbenchmarks it's often visible. We might be able to optimize away more of forEach 's overhead in the future, but in cases where you know that every CPU cycle matters, I recommend using plain old for (var i = 0; i < array.length; i++) loops.)