WIP: trying to understand benefits of "direct buffers"

WebReflection / coincident

An Atomics based Proxy to simplify, and synchronize, Worker related tasks.

MIT License

201 stars 3 forks source link

WIP: trying to understand benefits of "direct buffers" #33

Closed WebReflection closed 1 year ago

WebReflection commented 1 year ago

This MR tries to address #31 but:

to not lose the kind of typed view, it's still necessary to convert the typed array into a string somehow
because the SharedArrayBuffer with Atomics works only with an Int32Array or BigInt64Array, there's no way to pollute directly the SAB with generic typed views
accordingly, as everything boils down to a string, and charCodeAt and TextDecoder.decode are still needed, it's unclear how much memory or performance boost one can gain by branching logic and bypassing serialization

I might just also quick check both times and see if it's substantially faster to do this way, but I think I am going to close this and flag the related issue as won't fix.

WebReflection commented 1 year ago

this is with the current PR:

serialize: 2.750927734375 ms
es.js:2 deserialize: 1.40693359375 ms

this is with the current module as is, without this PR:

serialize: 4.90185546875 ms
es.js:2 deserialize: 3.505126953125 ms

in short, passing huge buffers is roughly twice as fast in here, but as we're speaking ms I am not sure I want to move forward with this idea, but it's OK to have already experimented with it in case I'll change my mind in the future.