.write(Buffer) support #174

AVVS · 2023-05-16T05:22:51Z

mashed up quick support for .write(chunk: Buffer), works rather well as long as input is a Buffer, in case it is a utf8 string performance degradation is quite bad. See #173 for why it might be needed

if raw buffer is something that is worth exploring further then I'd be happy to put more effort into having fast paths for both types of encoding, probably by preselecting mode of operation at the time of SonicBoom instantiation

benchSonic*1000: 2.189s
benchSonicSync*1000: 8.190s
benchSonic4k*1000: 2.147s
benchSonicSync4k*1000: 1.827s
benchCore*1000: 2.561s
benchConsole*1000: 4.835s
benchSonicBuf*1000: 840.388ms
benchSonicSyncBuf*1000: 6.882s
benchSonic4kBuf*1000: 824.605ms
benchSonicSync4kBuf*1000: 481.976ms
benchCoreBuf*1000: 824.466ms
benchSonic*1000: 2.105s
benchSonicSync*1000: 8.275s
benchSonic4k*1000: 2.179s
benchSonicSync4k*1000: 1.811s
benchCore*1000: 2.220s
benchConsole*1000: 5.287s
benchSonicBuf*1000: 856.635ms
benchSonicSyncBuf*1000: 6.852s
benchSonic4kBuf*1000: 849.188ms
benchSonicSync4kBuf*1000: 480.347ms
benchCoreBuf*1000: 802.708ms

mcollina

You have found why this supports only strings.

I think supporting buffers might be helpful but I don't think it's worthwhile - you might as well use Node.js core fs streams instead. I don't expect any solution would outperform them by a significant margin.

mcollina · 2023-05-16T09:26:42Z

index.js

+    return bufs[0]
+  }
+
+  return Buffer.concat(bufs, len)


This is the source of the slowdown. We should not merge them, but rather keep them as a list.

I've tried having a list here, but tests that expect single flush instead of multiples break. Will work on it further now that there are 2 separate content modes

All in all Buffer.concat seem to be cheaper than multiple fs.write as long as it doesn't have to allocate memory outside of the Buffer.poolSize. Will investigate further how that could be avoided

tried fs.writev to avoid concat, but that made it slower than writeStream

AVVS · 2023-05-16T19:01:16Z

benchSonicBuf*1000: 595.338ms

benchSonic4kBuf*1000: 590.645ms

benchCoreBuf*1000: 1.073s

There is at least a 2x speed up on my system with this vs core streams when writing buffers, plus proper error handling, so might be worth it, I'll see what the perf without merging small buffers is and whether it's possible to keep the exact performance profile for utf8

AVVS · 2023-05-16T19:56:19Z

Added contentMode setting. Not sure if the approach I've selected is the best from maintenance standpoint, but benchmarks are ok

benchSonic*1000: 428.597ms
benchSonicSync*1000: 6.054s
benchSonic4k*1000: 408.553ms
benchSonicSync4k*1000: 240.762ms

benchCore*1000: 2.785s
benchConsole*1000: 4.752s

benchSonicBuf*1000: 468.216ms
benchSonicSyncBuf*1000: 6.271s
benchSonic4kBuf*1000: 458.785ms
benchSonicSync4kBuf*1000: 343.907ms
benchCoreBuf*1000: 1.081s

benchSonic*1000: 407.736ms
benchSonicSync*1000: 6.123s
benchSonic4k*1000: 403.239ms
benchSonicSync4k*1000: 236.555ms

benchCore*1000: 2.768s
benchConsole*1000: 5.307s

benchSonicBuf*1000: 455.687ms
benchSonicSyncBuf*1000: 6.225s
benchSonic4kBuf*1000: 454.69ms
benchSonicSync4kBuf*1000: 334.973ms
benchCoreBuf*1000: 1.079s

AVVS · 2023-05-16T19:58:23Z

Master branch perf for comparison, so utf8 mode (default one) seem to be unaffected by changes

benchSonic*1000: 452.573ms
benchSonicSync*1000: 6.064s
benchSonic4k*1000: 429.246ms
benchSonicSync4k*1000: 258.853ms
benchCore*1000: 2.722s
benchConsole*1000: 4.725s

benchSonic*1000: 422.106ms
benchSonicSync*1000: 6.054s
benchSonic4k*1000: 426.696ms
benchSonicSync4k*1000: 253.348ms
benchCore*1000: 2.724s
benchConsole*1000: 5.305s

Only downside I see is that I have copied non-buffer related functions and changed logic there, so if any logic changes in the future then there are multiple places where this needs to be addressed

mcollina · 2023-05-17T08:59:10Z

@mmarchini could you take a look as well?

mmarchini · 2023-06-18T18:33:11Z

Sorry, was drowned in GitHub Notifications for months. I'll take a look at it this week.

mmarchini

overall looks fine, I didn't review the tests yet. Main concern as you mentioned is code duplication, so it'll be up to the project to decide if taking the burden of having some logic implemented in two places is worth it for this feature.

As for the performance aspects of this, I can see concat being faster than multiple writes, although that'll be highly dependent on system specs. Can you share specs for the machine where you ran your benchmarks?

mmarchini · 2023-06-20T19:02:14Z

index.js

@@ -56,7 +57,11 @@ function openFile (file, sonic) {

    // start
    if (!sonic._writing && sonic._len > sonic.minLength && !sonic.destroyed) {
-      actualWrite(sonic)
+      if (sonic.contentMode === 'buffer') {


Wondering if contentMode as an enum instead of a string would make more sense. Either way, this should be included in types/index.d.ts and should be documented.

great idea!

mmarchini · 2023-06-20T19:02:55Z

index.js

+        actualWriteBuffer(sonic)
+      } else {
+        actualWrite(sonic)
+      }


Can we call _actualWrite instead of having an if here?

if we assign it as sonic._actualWrite I think it's possible to do so. At the same time these are not hot paths, so not sure if it's worth it

Oh, I was thinking of using _actualWrite for maintainability rather than performance (V8 would take care of ensuring these ifs are performant so I'm not worried about that in from a perf standpoint)

mmarchini · 2023-06-20T19:15:56Z

index.js

+      actualWriteBuffer(this)
+    } else {
+      actualWrite(this)
+    }


same here, can't we call _actualWrite

AVVS · 2023-06-20T23:03:14Z

overall looks fine, I didn't review the tests yet. Main concern as you mentioned is code duplication, so it'll be up to the project to decide if taking the burden of having some logic implemented in two places is worth it for this feature.

As for the performance aspects of this, I can see concat being faster than multiple writes, although that'll be highly dependent on system specs. Can you share specs for the machine where you ran your benchmarks?

Chip Apple M1 Max
Memory 64 GB

AVVS · 2023-06-29T16:21:53Z

Updates incoming, in terms of bench

Node 20 performs ~25% better, Node 18 ~ 10%

node 20.3.1

benchSonic*1000: 411.93ms
benchSonicSync*1000: 6.231s
benchSonic4k*1000: 403.094ms
benchSonicSync4k*1000: 231.349ms
benchCore*1000: 1.900s
benchConsole*1000: 6.654s
benchSonicBuf*1000: 479.036ms
benchSonicSyncBuf*1000: 6.643s
benchSonic4kBuf*1000: 479.021ms
benchSonicSync4kBuf*1000: 338.352ms
benchCoreBuf*1000: 647.737ms
benchSonic*1000: 402.104ms
benchSonicSync*1000: 6.381s
benchSonic4k*1000: 411.276ms
benchSonicSync4k*1000: 230.133ms
benchCore*1000: 1.879s
benchConsole*1000: 7.073s
benchSonicBuf*1000: 488.614ms
benchSonicSyncBuf*1000: 6.739s
benchSonic4kBuf*1000: 468.49ms
benchSonicSync4kBuf*1000: 346.683ms
benchCoreBuf*1000: 657.002ms

node 18.16.0

benchSonic*1000: 435.906ms
benchSonicSync*1000: 6.272s
benchSonic4k*1000: 437.36ms
benchSonicSync4k*1000: 252.272ms
benchCore*1000: 2.252s
benchConsole*1000: 4.260s
benchSonicBuf*1000: 799.471ms
benchSonicSyncBuf*1000: 6.771s
benchSonic4kBuf*1000: 796.44ms
benchSonicSync4kBuf*1000: 645.533ms
benchCoreBuf*1000: 893.629ms
benchSonic*1000: 418.545ms
benchSonicSync*1000: 6.380s
benchSonic4k*1000: 417.882ms
benchSonicSync4k*1000: 245.55ms
benchCore*1000: 2.214s
benchConsole*1000: 5.374s
benchSonicBuf*1000: 787.246ms
benchSonicSyncBuf*1000: 6.849s
benchSonic4kBuf*1000: 822ms
benchSonicSync4kBuf*1000: 638.927ms
benchCoreBuf*1000: 888.464ms

AVVS · 2023-06-29T16:41:27Z

@mmarchini please take a look if the changes are sufficient based on your comments

AVVS · 2023-06-29T16:42:33Z

index.js

@@ -106,12 +110,35 @@ function SonicBoom (opts) {
  this.maxLength = maxLength || 0
  this.maxWrite = maxWrite || MAX_WRITE
  this.sync = sync || false
+  this.writable = true


forgot to mention, this is so that node.js stream util compose function works properly with this

mcollina · 2023-06-29T18:45:38Z

wow, quite a speedup also on writing things compared to core.

AVVS · 2023-06-29T18:49:36Z

Would be interesting to bench this with io_uring support as fs operations should be much faster on linux

Some weirdness is going on with win/node-14 combo during dep install

mcollina · 2023-06-29T18:51:39Z

Some weirdness is going on with win/node-14 combo during dep install

GHA is broken on that platform, just exclude it in the workflow file.

mmarchini · 2023-06-30T03:59:25Z

I'll try to take a look over the weekend :)

Would be interesting to bench this with io_uring support as fs operations should be much faster on linux

I have a Linux machine I can try it on. Will report back with the results

AVVS · 2023-08-04T16:13:06Z

anything I should do with this to move it forward? :)

mcollina · 2023-09-12T07:32:49Z

@mmarchini could you take a quick look? I would prefer if this did not break you. Otherwise I'm ok to ship it in the coming weeks.

mcollina

lgtm

AVVS added 3 commits May 15, 2023 21:52

feat: handle buffer or utf8

9152450

chore: adjust bench to show raw buffer perf

1ed75f0

chore: empty out lengths array

99b0470

mcollina requested changes May 16, 2023

View reviewed changes

feat: contentMode setting

a694b87

mmarchini reviewed Jun 20, 2023

View reviewed changes

chore: contentMode in .d.ts, save _actualWrite

ad82abc

AVVS commented Jun 29, 2023

View reviewed changes

chore: exclude win+node14

7b1079f

jsumners requested review from mmarchini and mcollina August 28, 2023 16:22

mcollina approved these changes Oct 3, 2023

View reviewed changes

Merge branch 'master' into feat/buf-support

5764633

mcollina changed the title ~~PoC: .write(Buffer) support, related to #173~~ .write(Buffer) support Oct 3, 2023

mcollina merged commit 2764745 into pinojs:master Oct 3, 2023
10 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

.write(Buffer) support #174

.write(Buffer) support #174

AVVS commented May 16, 2023 •

edited

Loading

mcollina left a comment

mcollina May 16, 2023

AVVS May 16, 2023

AVVS May 16, 2023

AVVS May 16, 2023

AVVS commented May 16, 2023

AVVS commented May 16, 2023

AVVS commented May 16, 2023

mcollina commented May 17, 2023

mmarchini commented Jun 18, 2023

mmarchini left a comment

mmarchini Jun 20, 2023

AVVS Jun 20, 2023

mmarchini Jun 20, 2023

AVVS Jun 20, 2023

mmarchini Jun 21, 2023 •

edited

Loading

mmarchini Jun 20, 2023

AVVS commented Jun 20, 2023

AVVS commented Jun 29, 2023

AVVS commented Jun 29, 2023

AVVS Jun 29, 2023

mcollina commented Jun 29, 2023

AVVS commented Jun 29, 2023

mcollina commented Jun 29, 2023

mmarchini commented Jun 30, 2023

AVVS commented Aug 4, 2023

mcollina commented Sep 12, 2023

mcollina left a comment

.write(Buffer) support #174

.write(Buffer) support #174

Conversation

AVVS commented May 16, 2023 • edited Loading

mcollina left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

AVVS commented May 16, 2023

AVVS commented May 16, 2023

AVVS commented May 16, 2023

mcollina commented May 17, 2023

mmarchini commented Jun 18, 2023

mmarchini left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

mmarchini Jun 21, 2023 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

AVVS commented Jun 20, 2023

AVVS commented Jun 29, 2023

AVVS commented Jun 29, 2023

Choose a reason for hiding this comment

mcollina commented Jun 29, 2023

AVVS commented Jun 29, 2023

mcollina commented Jun 29, 2023

mmarchini commented Jun 30, 2023

AVVS commented Aug 4, 2023

mcollina commented Sep 12, 2023

mcollina left a comment

Choose a reason for hiding this comment

AVVS commented May 16, 2023 •

edited

Loading

mmarchini Jun 21, 2023 •

edited

Loading