Node: v9.x seems to be slowing down ?

[Lizhooh]

v9.x seems to be slowing down?
I have a computational function that performs benchmark tests in different versions of node.

// test function
function compute(w, v, s, k, f, vr1 = 1, vr2 = 0) {
    return s + (Math.max(v * vr2 || v / vr1, w) - f) * k;
}

Results are as follows:

(nodejs v7.9) 63,482,006 ops/sec ±0.93% (90 runs sampled)
(nodejs v8.7) 539,999,198 ops/sec ±0.48% (89 runs sampled)
(nodejs v8.9) 543,188,459 ops/sec ±0.53% (91 runs sampled)
(nodejs v9.1) 487,698,977 ops/sec ±0.64% (89 runs sampled)

[vsemozhetbyt]

Hi, @Lizhooh! Thank you for the information. Could you cite the complete runnable benchmark code?

cc @nodejs/performance, @nodejs/v8.

[Lizhooh]

@vsemozhetbyt

const Benchmark = require('benchmark');
const suite = new Benchmark.Suite;

// test function
function compute(w, v, s, k, f, vr1 = 1, vr2 = 0) {
    return s + (Math.max(v * vr2 || v / vr1, w) - f) * k;
}

suite
    .add('my compute', function () {
        let result = compute(32, 27 / 1e6, 20, 3, 10, 1, 347);
    })
    .on('cycle', function (event) {
        console.log(String(event.target));
    })
    .run({ 'async': true });

[vsemozhetbyt]

@Lizhooh Thank you. It seems this pattern does slow down even more with the next V8 versions. My results on Windows 7 x64:

Node.js 8.9.1 (V8 6.1)
my compute x 271,200,039 ops/sec ±0.16% (100 runs sampled)

Node.js 9.2.0 (V8 6.2)
my compute x 259,525,541 ops/sec ±0.73% (91 runs sampled)

Node.js 9.0.0 (V8 6.3) canary
my compute x 132,483,285 ops/sec ±0.16% (104 runs sampled)

Node.js 10.0.0 (V8 6.4) canary
my compute x 132,412,275 ops/sec ±0.22% (105 runs sampled)

I hope @nodejs/v8 will chime in with some info.

[peterwmwong]

The perf difference appears to be mostly coming from the calculation for the first argument of Math.max(), v * vr2 || v / vr1.

I think I've reduced the repro to the following:

'use strict';

function run(i) {
  const v = 0.1; // Important #1: Must be < 1.0 (>= 1.0 seems fine)
  return (
    (v * 1.0) || // Important #2: Can be any value, but must perform some operation like add/sub/mult/divide
    (v * 1.0)    // Important #3: Can be any value, but must perform some operation like add/sub/mult/divide
  );
}

function bench(iters) {
  let result = 0;
  for (let iter = 0; iter < iters; iter++) {
    result += run(iter);
  }
  return result;
}

const start = performance.now();
bench(1e7);
const end = performance.now();
console.log(end - start);

Comparing 6.1-lkgr fe8c059def and master aae52f6b0f

| | Before | After | Change |
|--------|--------|-------|-------------|
| | 12.42ms | 45.85ms | 0.27x |

[benjamingr]

@bmeurer

[bmeurer]

Thanks for the awesome repro. It seems that the loop peeling doesn't properly peel the loop in the bench function anymore.

Upstream issue: v8:7099

[YurySolovyov]

Link is broken

[vsemozhetbyt]

Maybe https://bugs.chromium.org/p/v8/issues/detail?id=7099

[apapirovski]

Not much left to do here.