Hacker News Reader

Comment by cogman10 | Hacker News Reader

Too bad that's a flawed benchmarking methodolgy. JITs are notoriously hard to correctly profile and the benchmark lib isn't even sort of doing the right thing.

For example, it's missing warmup. The results aren't being consumed in a way that wouldn't optimize them away. The framework itself imposes a pretty large amount of overhead (moreso than I'd expect from leftpad).

It is somewhat likely that what they are measuring isn't leftpad performance, but rather how fast the JIT ends up optimizing the benchmark code.

I'd suggest watching this video https://vimeo.com/78900556

It's about microprofiling the JVM, but the principles are the same for other JIT based languages (such as javascript).

View on HN

Yeah benchmarks can be much more difficult to get right than it might seem at first glance. Good thing they didn't try to write their own benchmarking code, otherwise they might have fallen into those traps you just mentioned.

Luckily, they didn't, and instead pulled in the the `benchmark` library as a development dependency. The author of said library works on V8, and already considered all those problems and much, much, more[1].

[1]: https://mathiasbynens.be/notes/javascript-benchmarking

You are making the assumption that the library doesn't have the flaws I mentioned. It does (You can go read the source yourself https://raw.githubusercontent.com/bestiejs/benchmark.js/2.1.... )

There's no portion of the code that does warmups. There's no portion of code that "blackholes" the results to keep the JIT from optimizing away the code under benchmark. There is a lot of code though... so that's... good?

You make the assumption that just because a lib is popular or widely used it is "correct" or "the best". When it comes to microbenchmarks, that's usually flawed. Very VERY few people actually get them right, benchmark.js is no exception.

That, of course, doesn't mean that benchmark.js can't be useful. For macrobenchmarks it will be roughly right. However, for something as small as leftpad, it's almost certainly not the right way to measure performance.

Okay, you win. I'm not going to read the entire source of that package just to make a point. Though I do find it strange that the author of that library would write an entire blog post on the topic and then not take his own advice in the implementation of the library he wrote.

Tell me, where in that blog post does he mention doing warm up cycles or avoiding having the JIT optimize away the method? (Hint, he doesn't mention that... so, no, he didn't actually miss his own advice.)

The article is completely consumed with getting the timing of benchmarking right. Which, to be fair, is a place where microbenchmarks often go wrong. It, however, isn't the ONLY place they go wrong.

> There's no portion of the code that does warmups.

This isn't true - Benchmark.js will repeatedly rerun benchmarks until it gathers statistically meaningful results.

> There's no portion of code that "blackholes" the results to keep the JIT from optimizing away the code under benchmark.

True, and there's actually nothing benchmark.js can do to ensure that doesn't happen in the general case but when this does happen the results are usually pretty obvious - we'd see billions of ops/sec. Incidentally the left-pad benchmarks do not suffer from this issue.

> Luckily, they didn't, and instead pulled in the the `benchmark` library as a development dependency. The author of said library works on V8, and already considered all those problems and much, much, more[1].

This is the "benchmark lib" that was mentioned in the very first sentece of the comment you replied to.

Reply on news.ycombinator.com