stream: replace queue and readRequests with buffer list for performance improvement #49554

rluvaton · 2023-09-08T12:12:42Z

webstream possible improvement:

`BufferList` implementation

00:24:29.343                                                                        confidence improvement accuracy (*)   (**)   (***)
00:24:29.343 webstreams/creation.js kind='ReadableStream' n=50000                                   1.04 %       ±3.94% ±5.25%  ±6.83%
00:24:29.343 webstreams/creation.js kind='TransformStream' n=50000                                  0.06 %       ±2.59% ±3.45%  ±4.52%
00:24:29.343 webstreams/creation.js kind='WritableStream' n=50000                                   0.19 %       ±4.60% ±6.12%  ±7.98%
00:24:29.343 webstreams/pipe-to.js highWaterMarkW=1024 highWaterMarkR=1024 n=100000                 1.58 %       ±4.24% ±5.64%  ±7.34%
00:24:29.343 webstreams/pipe-to.js highWaterMarkW=1024 highWaterMarkR=2048 n=100000                 1.13 %       ±4.28% ±5.69%  ±7.41%
00:24:29.343 webstreams/pipe-to.js highWaterMarkW=1024 highWaterMarkR=4096 n=100000                 1.50 %       ±4.50% ±5.99%  ±7.79%
00:24:29.343 webstreams/pipe-to.js highWaterMarkW=1024 highWaterMarkR=512 n=100000                  1.04 %       ±5.26% ±7.00%  ±9.12%
00:24:29.343 webstreams/pipe-to.js highWaterMarkW=2048 highWaterMarkR=1024 n=100000                 2.24 %       ±5.31% ±7.07%  ±9.20%
00:24:29.343 webstreams/pipe-to.js highWaterMarkW=2048 highWaterMarkR=2048 n=100000                 3.03 %       ±4.62% ±6.15%  ±8.00%
00:24:29.343 webstreams/pipe-to.js highWaterMarkW=2048 highWaterMarkR=4096 n=100000                 2.97 %       ±5.12% ±6.81%  ±8.86%
00:24:29.343 webstreams/pipe-to.js highWaterMarkW=2048 highWaterMarkR=512 n=100000                  3.92 %       ±4.65% ±6.19%  ±8.06%
00:24:29.343 webstreams/pipe-to.js highWaterMarkW=4096 highWaterMarkR=1024 n=100000                -0.04 %       ±4.14% ±5.50%  ±7.16%
00:24:29.343 webstreams/pipe-to.js highWaterMarkW=4096 highWaterMarkR=2048 n=100000                -1.05 %       ±4.87% ±6.48%  ±8.43%
00:24:29.343 webstreams/pipe-to.js highWaterMarkW=4096 highWaterMarkR=4096 n=100000                 2.22 %       ±4.57% ±6.07%  ±7.91%
00:24:29.343 webstreams/pipe-to.js highWaterMarkW=4096 highWaterMarkR=512 n=100000                 -2.14 %       ±4.71% ±6.27%  ±8.16%
00:24:29.344 webstreams/pipe-to.js highWaterMarkW=512 highWaterMarkR=1024 n=100000                  1.56 %       ±4.40% ±5.85%  ±7.62%
00:24:29.344 webstreams/pipe-to.js highWaterMarkW=512 highWaterMarkR=2048 n=100000                  0.54 %       ±4.56% ±6.06%  ±7.89%
00:24:29.344 webstreams/pipe-to.js highWaterMarkW=512 highWaterMarkR=4096 n=100000                  3.16 %       ±4.44% ±5.91%  ±7.70%
00:24:29.344 webstreams/pipe-to.js highWaterMarkW=512 highWaterMarkR=512 n=100000                   0.90 %       ±4.70% ±6.26%  ±8.16%
00:24:29.344 webstreams/readable-async-iterator.js n=100000                                         0.03 %       ±7.40% ±9.84% ±12.81%
00:24:29.344 
00:24:29.344 Be aware that when doing many comparisons the risk of a false-positive
00:24:29.344 result increases. In this case, there are 20 comparisons, you can thus
00:24:29.344 expect the following amount of false-positive results:
00:24:29.344   1.00 false positives, when considering a   5% risk acceptance (*, **, ***),
00:24:29.344   0.20 false positives, when considering a   1% risk acceptance (**, ***),
00:24:29.344   0.02 false positives, when considering a 0.1% risk acceptance (***)

From my local tests:

Before

❯ ./node ./benchmark/webstreams/pipe-to.js
webstreams/pipe-to.js highWaterMarkW=512 highWaterMarkR=512 n=5000000: 1,073,259.5235043082
webstreams/pipe-to.js highWaterMarkW=1024 highWaterMarkR=512 n=5000000: 1,112,582.388083188
webstreams/pipe-to.js highWaterMarkW=2048 highWaterMarkR=512 n=5000000: 1,130,721.9334553012
webstreams/pipe-to.js highWaterMarkW=4096 highWaterMarkR=512 n=5000000: 1,137,875.2107408328
webstreams/pipe-to.js highWaterMarkW=512 highWaterMarkR=1024 n=5000000: 1,145,415.4228369598
webstreams/pipe-to.js highWaterMarkW=1024 highWaterMarkR=1024 n=5000000: 1,119,071.6166112064
webstreams/pipe-to.js highWaterMarkW=2048 highWaterMarkR=1024 n=5000000: 1,112,401.620591177
webstreams/pipe-to.js highWaterMarkW=4096 highWaterMarkR=1024 n=5000000: 1,122,177.222643778
webstreams/pipe-to.js highWaterMarkW=512 highWaterMarkR=2048 n=5000000: 1,154,120.0021085772
webstreams/pipe-to.js highWaterMarkW=1024 highWaterMarkR=2048 n=5000000: 1,137,931.8271974607
webstreams/pipe-to.js highWaterMarkW=2048 highWaterMarkR=2048 n=5000000: 1,150,085.2975136968
webstreams/pipe-to.js highWaterMarkW=4096 highWaterMarkR=2048 n=5000000: 1,120,964.4298001872
webstreams/pipe-to.js highWaterMarkW=512 highWaterMarkR=4096 n=5000000: 1,126,744.873505527
webstreams/pipe-to.js highWaterMarkW=1024 highWaterMarkR=4096 n=5000000: 1,148,843.5697545216
webstreams/pipe-to.js highWaterMarkW=2048 highWaterMarkR=4096 n=5000000: 1,133,865.2778680783
webstreams/pipe-to.js highWaterMarkW=4096 highWaterMarkR=4096 n=5000000: 1,132,648.4636818648

After

❯ ./node ./benchmark/webstreams/pipe-to.js
webstreams/pipe-to.js highWaterMarkW=512 highWaterMarkR=512 n=5000000: 1,191,920.5427624737
webstreams/pipe-to.js highWaterMarkW=1024 highWaterMarkR=512 n=5000000: 1,199,980.044139869
webstreams/pipe-to.js highWaterMarkW=2048 highWaterMarkR=512 n=5000000: 1,196,830.8757150145
webstreams/pipe-to.js highWaterMarkW=4096 highWaterMarkR=512 n=5000000: 1,209,921.9163751516
webstreams/pipe-to.js highWaterMarkW=512 highWaterMarkR=1024 n=5000000: 1,199,785.7782492938
webstreams/pipe-to.js highWaterMarkW=1024 highWaterMarkR=1024 n=5000000: 1,191,548.2411812833
webstreams/pipe-to.js highWaterMarkW=2048 highWaterMarkR=1024 n=5000000: 1,205,993.3892632995
webstreams/pipe-to.js highWaterMarkW=4096 highWaterMarkR=1024 n=5000000: 1,191,874.810445273
webstreams/pipe-to.js highWaterMarkW=512 highWaterMarkR=2048 n=5000000: 1,209,462.7154360572
webstreams/pipe-to.js highWaterMarkW=1024 highWaterMarkR=2048 n=5000000: 1,187,768.408534668
webstreams/pipe-to.js highWaterMarkW=2048 highWaterMarkR=2048 n=5000000: 1,195,187.8509158778
webstreams/pipe-to.js highWaterMarkW=4096 highWaterMarkR=2048 n=5000000: 1,201,572.702366821
webstreams/pipe-to.js highWaterMarkW=512 highWaterMarkR=4096 n=5000000: 1,197,150.4825638772
webstreams/pipe-to.js highWaterMarkW=1024 highWaterMarkR=4096 n=5000000: 1,179,266.3281946357
webstreams/pipe-to.js highWaterMarkW=2048 highWaterMarkR=4096 n=5000000: 1,205,445.1767767125
webstreams/pipe-to.js highWaterMarkW=4096 highWaterMarkR=4096 n=5000000: 1,235,058.8880696953

ronag

I think using FixedQueue would have better performance?

rluvaton · 2023-09-08T12:27:06Z

lib/internal/webstreams/readablestream.js

+  // TODO - may be able to change this to next next next
  while (reader[kState].readRequests.length > 0) {
+    // TODO - why this is in the while loop?


remove todos

rluvaton · 2023-09-08T12:59:19Z

Benchmark CI for BufferList implementation

https://ci.nodejs.org/view/Node.js%20benchmark/job/benchmark-node-micro-benchmarks/1369/

rluvaton · 2023-09-08T14:10:33Z

the pipe-to benchmark takes too long, it's currently running for an hour.

Each iteration (single combination) takes 9 seconds. there are 16 possible combinations.
so 9 seconds * 16 combination * 30 runs * 2 versions = 2 hours and 24 minutes...

new Benchmark CI that has reduced the number of chunks from 5M to 100K for BufferList implementation
https://ci.nodejs.org/view/Node.js%20benchmark/job/benchmark-node-micro-benchmarks/1370/

Each iteration (single combination) takes 9 seconds. there are 16 possible combinations. so 9 seconds * 16 combination * 30 runs * 2 versions = 2 hours and 24 minutes...

rluvaton · 2023-09-08T14:39:57Z

lib/internal/webstreams/util.js

  controller[kState].queueTotalSize = 0;
 }

 function peekQueueValue(controller) {
  assert(controller[kState].queue !== undefined);
  assert(controller[kState].queueTotalSize !== undefined);
  assert(controller[kState].queue.length);
-  return controller[kState].queue[0].value;
+  debugger;


rluvaton · 2023-09-08T15:28:49Z

I think using FixedQueue would have better performance?

Could not make it work, the tests failed for some reason...

current BufferList does not show any improvement so closing this

nodejs-github-bot added needs-ci PRs that need a full CI run. web streams labels Sep 8, 2023

rluvaton force-pushed the improve-readable-stream-perf-by-moving-to-list branch from 8465d8b to 4a982c6 Compare September 8, 2023 12:15

rluvaton added performance Issues and PRs related to the performance of Node.js. needs-benchmark-ci PR that need a benchmark CI run. labels Sep 8, 2023

rluvaton changed the title ~~stream: replace queue with buffer list for performance improvement~~ stream: replace queue and readRequests with buffer list for performance improvement Sep 8, 2023

ronag reviewed Sep 8, 2023

View reviewed changes

rluvaton commented Sep 8, 2023

View reviewed changes

rluvaton added 2 commits September 8, 2023 16:59

stream: replace queue with buffer list for performance improvement

240542c

format

4e39d86

rluvaton force-pushed the improve-readable-stream-perf-by-moving-to-list branch from 5943045 to 4e39d86 Compare September 8, 2023 13:59

reduce number of chunks in pipe benchmark

04f02a1

Each iteration (single combination) takes 9 seconds. there are 16 possible combinations. so 9 seconds * 16 combination * 30 runs * 2 versions = 2 hours and 24 minutes...

rluvaton commented Sep 8, 2023

View reviewed changes

rluvaton closed this Sep 8, 2023

rluvaton deleted the improve-readable-stream-perf-by-moving-to-list branch September 8, 2023 15:28

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

stream: replace queue and readRequests with buffer list for performance improvement #49554

stream: replace queue and readRequests with buffer list for performance improvement #49554

rluvaton commented Sep 8, 2023 •

edited

Loading

ronag left a comment

rluvaton Sep 8, 2023

rluvaton commented Sep 8, 2023 •

edited

Loading

rluvaton commented Sep 8, 2023 •

edited

Loading

rluvaton Sep 8, 2023

rluvaton commented Sep 8, 2023

stream: replace queue and readRequests with buffer list for performance improvement #49554

stream: replace queue and readRequests with buffer list for performance improvement #49554

Conversation

rluvaton commented Sep 8, 2023 • edited Loading

BufferList implementation

ronag left a comment

Choose a reason for hiding this comment

rluvaton Sep 8, 2023

Choose a reason for hiding this comment

rluvaton commented Sep 8, 2023 • edited Loading

rluvaton commented Sep 8, 2023 • edited Loading

rluvaton Sep 8, 2023

Choose a reason for hiding this comment

rluvaton commented Sep 8, 2023

rluvaton commented Sep 8, 2023 •

edited

Loading

`BufferList` implementation

rluvaton commented Sep 8, 2023 •

edited

Loading

rluvaton commented Sep 8, 2023 •

edited

Loading