V8 memory quiz | Hacker News Reader

V8 memory quiz

mattzeunert.com

87 pointsmostlystatic11 comments

Spoiler: V8's default array implementation holds pointers to boxed objects [Questions 5 & 6], but there are special-case implementations for arrays of doubles [1] and 16-bit integers [9] (and presumably others). However, adding a single string to an array of doubles violates the special-case and requires every double to be boxed, significantly increasing memory footprint [2].

It's interesting to compare this to a "NaN-boxing" approach, which allows pointers and doubles (and other simple values) to all exist in a single "namespace" of 8-byte values. In the (uncommon?) case of a heterogeneous array, this can save memory by avoiding the need to box simple values. The downside is that NaN-boxing prevents the use of full 64-bit pointers.

This isn't a special cased array specialization for 16-bit integers (16-bit integers are so rare that special casing those is not worth it). Note that a standard array is not used in [9], rather an `Int16Array` is being used. Due to JavaScript specification requirements (arrays of different integer sizes can share storage), the ONLY sane implementation of Int16Array is well... an array of 16-bit integers.

This is one of my favourite ways to play with a concept. The quiz part forces my brain to turn on the problem solving section rather than just the reading section.

Okay, but how about other JavaScript engines?

And shouldn't stuff like this be in the specs, so that the language becomes more consistent and predictable?

> Okay, but how about other JavaScript engines?

SpiderMonkey uses NaN-boxing, so pointers to objects will sometimes take up less space than in V8. As for hidden classes, I'm not sure about how shapes (SpiderMonkey's term for hidden classes) interact with constructors.

> And shouldn't stuff like this be in the specs, so that the language becomes more consistent and predictable?

That would stifle compiler innovation. Not even the "low-level", "predictable" C and C++ mandate particular implementations, beyond specifying the bare minimum necessary to ensure pointer arithmetic works portably.

> Not even the "low-level", "predictable" C and C++ mandate particular implementations

But at least with C and C++ a developer can test the performance before deployment.

These are implementation details -- they don't belong in the specs. You can't access raw memory from JavaScript.

No, since you can obviously implement Javascript in ways that have other memory usage patterns.

So V8 doesn't use compressed pointers? It sounds wasteful to use whole 8 bytes for pointer if typical JS application is unlikely to need more than 2GB of ram

The default size of an array for V8 is over 1304209 elements?

When it runs out of space, it resizes to (at least) a multiple of its current space to avoid having to resize too often, making pushing a constant-time operation on average. 1304209 is the smallest size in the sequence starting with an empty array that’s at least 1048576.

Reply on news.ycombinator.com