Repository navigation
Blob.text and Blob.arraybuffer are 10x+ slower on 20.x #118
Description
Activity
Some benchmark data:
NodeJS 16.20.1
name ops/sec samples text (128) 17,269 57 text (1024) 10,257 57 arrayBuffer (128) 24,974 53 arrayBuffer (1024) 10,955 57 slice (0, 64) 48,310 67 slice (0, 512) 23,750 66 NodeJS 18.17.0
name ops/sec samples text (128) 36,176 74 text (1024) 13,691 69 arrayBuffer (128) 36,788 74 arrayBuffer (1024) 14,305 75 slice (0, 64) 49,139 71 slice (0, 512) 26,774 76 NodeJS 20.5.0
name ops/sec samples text (128) 3,194 74 text (1024) 380 76 arrayBuffer (128) 3,478 80 arrayBuffer (1024) 407 75 slice (0, 64) 45,076 66 slice (0, 512) 16,484 64 Reacted by Fabian Meyer, Benjamin Gruenbaum, Carlos Fuentes, Raz Luvaton and Debadree Chatterjee- changed the title
[-]Blob can be faster[/-][+]Blob.text and Blob.arraybuffer are 10x+ slower on 20.x[/+]on Sep 21, 2023 @H4ad run a git bisect to find where the regression is?
@benjamingr I will!
Also, this regression was so hard to run this piece of code:
const buff = Buffer.allocUnsafe(1024 ** 2); const source = new File(buff, 'dummy.txt', options); const now = performance.now(); source.text().then(() => console.log(`diff: ${performance.now() - now}`));
It took 3s on node 20.x and 80ms on 18.x.
Which public api affected by this?
If I understand your question,
Fileis fromBufferwhich inheritsBlob, so any code that inherits and usesBlobwill be slowdown when calling thetextorarraybuffer.From what I read,
undiciusesFileandBlobto handlemultipart/form-dataand files, so is very likely they had this regression too if they call.text.This will be a tricky one as the new mechanism underlying the C++ implementation of Blob is also going to be used for the QUIC implementation that I'm currently working on. It also serves as the underlying bit for file-backed
Blob. There is lots of room for performance improvement here but I would appreciate this not being changed right at this moment as it could lead to conflicts with the ongoing QUIC implementation work. What would probably be the most valuable right now would be just identifying where the key bottlenecks -- identifying those so that we can circle back around to making those pieces faster.Reacted by flakey5, Vinicius Lourenço, Yagiz Nizipli, Robert Nagy, Debadree Chatterjee and Carlos FuentesIn the QUIC implementation, are you changing the Blob?
- added a commit that references this issue
on Sep 26, 2023 Making an initial attempt here nodejs/node#49873 just stuck on how to make the pulling part async similar to how we are doing
queueMicrotaskin js in cpp, Is there maybe some sort of way to queue jobs in c++ part 🤔so I kinda did git bisect and I suspect it was either created by:
or created by those:
Reacted by Benjamin GruenbaumReacted by Vinicius LourençoHmm the main problem seems like we pass callback -> callback again passes back value after reading -> repeat 2 JS -> C++ and back calls everytime data being read
just stuck on how to make the pulling part async similar to how we are doing queueMicrotask in js in cpp, Is there maybe some sort of way to queue jobs in c++ part
Pretty sure you were looking for
v8::Isolate::EnqueueMicrotaskReacted by Debadree Chatterjee and Carlos Fuentesjust stuck on how to make the pulling part async similar to how we are doing queueMicrotask in js in cpp, Is there maybe some sort of way to queue jobs in c++ part
Pretty sure you were looking for
v8::Isolate::EnqueueMicrotaskahah yes! would do the job i guess but bigger problem rn is how to queue them described in nodejs/node#49873 (comment)
Made some progress on this in nodejs/node#49873 but stuck on some C++ pointer stuff if anyone could help or take a look would be great
- added a commit that references this issue
on Apr 15, 2024
The
blob/blob.jsbenchmark is taking hours: https://ci.nodejs.org/view/Node.js%20benchmark/job/benchmark-node-micro-benchmarks/1402/consoleThis probably didn't take so long in the past, there is this PR that changed the Blob's behavior a lot: nodejs/node#45258
From what I saw, the problem is in the JS->C++ transitions, for each piece of data it crosses JS->C++.
Furthermore, the
textfunction expectsarraybufferand then callsdecode, when we could justdecodethe data as soon as we get it from the C++ instead of keeping everything in memory and then discarding it.I can do some optimizations on the JS side but I think this could be largely improved if we do the heavy operations on C++, I have no idea how to handle the asynchronous code on C++, so I accept some hints or if anyone wants to try work on this.