Feature/writev master #5257

indutny · 2013-04-08T22:07:09Z

Previous discussion: #5246

indutny · 2013-04-09T07:27:27Z

So its like "mostly faster" on http benchmarks.

indutny · 2013-04-09T10:01:19Z

Updated pull request to be more consistent with newly added coalesce API for streams2. The latests commits are introducing tcp.writev method which can accept mixed array of buffers and strings with encodings, under the hood it copies them to one storage and invoke uv_write() only once. Benchmark results are coming soon...

indutny · 2013-04-09T10:16:35Z

Removed all benchmark results, just realized that I was running them on CI server.

indutny · 2013-04-09T13:44:13Z

Benchmark results (sorry difference is inversed):

http/chunked.js num=1 size=1 c=100: ./node-original: 1350.7 ./node: 1635.6 ........................... -17.42%
http/chunked.js num=1 size=64 c=100: ./node-original: 1338.4 ./node: 1617.9 .......................... -17.27%
http/chunked.js num=1 size=256 c=100: ./node-original: 1334.5 ./node: 1656.6 ......................... -19.44%
http/chunked.js num=4 size=1 c=100: ./node-original: 463.64 ./node: 754.69 ........................... -38.56%
http/chunked.js num=4 size=64 c=100: ./node-original: 472.62 ./node: 734.34 .......................... -35.64%
http/chunked.js num=4 size=256 c=100: ./node-original: 453 ./node: 725.06 ............................ -37.52%
http/chunked.js num=8 size=1 c=100: ./node-original: 247 ./node: 417.25 .............................. -40.80%
http/chunked.js num=8 size=64 c=100: ./node-original: 243.37 ./node: 415.04 .......................... -41.36%
http/chunked.js num=8 size=256 c=100: ./node-original: 238.49 ./node: 413.63 ......................... -42.34%
http/chunked.js num=16 size=1 c=100: ./node-original: 121.18 ./node: 229.34 .......................... -47.16%
http/chunked.js num=16 size=64 c=100: ./node-original: 123.02 ./node: 225.3 .......................... -45.40%
http/chunked.js num=16 size=256 c=100: ./node-original: 122 ./node: 227.8 ............................ -46.44%
http/cluster.js type=bytes length=4 c=50: ./node-original: 3285 ./node: 3194.8 ......................... 2.82%
http/cluster.js type=bytes length=4 c=500: ./node-original: 3200.1 ./node: 3168.4 ...................... 1.00%
http/cluster.js type=bytes length=1024 c=50: ./node-original: 3203 ./node: 3254 ....................... -1.57%
http/cluster.js type=bytes length=1024 c=500: ./node-original: 3178.3 ./node: 3107.4 ................... 2.28%
http/cluster.js type=bytes length=102400 c=50: ./node-original: 471.93 ./node: 470.87 .................. 0.23%
http/cluster.js type=bytes length=102400 c=500: ./node-original: 487.8 ./node: 487.2 ................... 0.12%
http/cluster.js type=buffer length=4 c=50: ./node-original: 3312.9 ./node: 3130.9 ...................... 5.81%
http/cluster.js type=buffer length=4 c=500: ./node-original: 3094.8 ./node: 3023.1 ..................... 2.37%
http/cluster.js type=buffer length=1024 c=50: ./node-original: 3096.6 ./node: 3295.8 .................. -6.04%
http/cluster.js type=buffer length=1024 c=500: ./node-original: 3107.8 ./node: 3010 .................... 3.25%
http/cluster.js type=buffer length=102400 c=50: ./node-original: 1694.8 ./node: 1681.3 ................. 0.81%
http/cluster.js type=buffer length=102400 c=500: ./node-original: 1877.5 ./node: 1792.1 ................ 4.77%
http/end-vs-write-end.js type=asc kb=64 c=100 method=write: ./node-original: 938.68 ./node: 1077.3 ... -12.87%
http/end-vs-write-end.js type=asc kb=64 c=100 method=end  : ./node-original: 1096 ./node: 1154.9 ...... -5.09%
http/end-vs-write-end.js type=asc kb=128 c=100 method=write: ./node-original: 656.37 ./node: 701.15 ... -6.39%
http/end-vs-write-end.js type=asc kb=128 c=100 method=end  : ./node-original: 716.93 ./node: 712.99 .... 0.55%
http/end-vs-write-end.js type=asc kb=256 c=100 method=write: ./node-original: 403.77 ./node: 368.3 ..... 9.63%
http/end-vs-write-end.js type=asc kb=256 c=100 method=end  : ./node-original: 420.78 ./node: 416.51 .... 1.03%
http/end-vs-write-end.js type=asc kb=1024 c=100 method=write: ./node-original: 103.39 ./node: 115.53 . -10.51%
http/end-vs-write-end.js type=asc kb=1024 c=100 method=end  : ./node-original: 101.72 ./node: 104.2 ... -2.38%
http/end-vs-write-end.js type=utf kb=64 c=100 method=write: ./node-original: 1117.3 ./node: 1286.8 ... -13.18%
http/end-vs-write-end.js type=utf kb=64 c=100 method=end  : ./node-original: 1330.9 ./node: 1360.6 .... -2.18%
http/end-vs-write-end.js type=utf kb=128 c=100 method=write: ./node-original: 739.31 ./node: 810.58 ... -8.79%
http/end-vs-write-end.js type=utf kb=128 c=100 method=end  : ./node-original: 824.67 ./node: 840.18 ... -1.85%
http/end-vs-write-end.js type=utf kb=256 c=100 method=write: ./node-original: 492.41 ./node: 531.61 ... -7.37%
http/end-vs-write-end.js type=utf kb=256 c=100 method=end  : ./node-original: 535.62 ./node: 508.7 ..... 5.29%
http/end-vs-write-end.js type=utf kb=1024 c=100 method=write: ./node-original: 163.93 ./node: 144.73 .. 13.26%
http/end-vs-write-end.js type=utf kb=1024 c=100 method=end  : ./node-original: 162.43 ./node: 162.9 ... -0.29%
http/end-vs-write-end.js type=buf kb=64 c=100 method=write: ./node-original: 1043.3 ./node: 1677.5 ... -37.81%
http/end-vs-write-end.js type=buf kb=64 c=100 method=end  : ./node-original: 1681.6 ./node: 1607.3 ..... 4.62%
http/end-vs-write-end.js type=buf kb=128 c=100 method=write: ./node-original: 872.07 ./node: 1481.3 .. -41.13%
http/end-vs-write-end.js type=buf kb=128 c=100 method=end  : ./node-original: 880.55 ./node: 1363.6 .. -35.42%
http/end-vs-write-end.js type=buf kb=256 c=100 method=write: ./node-original: 723.88 ./node: 1106.5 .. -34.58%
http/end-vs-write-end.js type=buf kb=256 c=100 method=end  : ./node-original: 746.71 ./node: 1111.7 .. -32.83%
http/end-vs-write-end.js type=buf kb=1024 c=100 method=write: ./node-original: 368.01 ./node: 492.63 . -25.30%
http/end-vs-write-end.js type=buf kb=1024 c=100 method=end  : ./node-original: 368.08 ./node: 509.32 . -27.73%
http/simple.js type=bytes length=4 c=50: ./node-original: 2204.9 ./node: 2268.5 ....................... -2.80%
http/simple.js type=bytes length=4 c=500: ./node-original: 2083.3 ./node: 2010 ......................... 3.64%
http/simple.js type=bytes length=1024 c=50: ./node-original: 2198.5 ./node: 2350.3 .................... -6.46%
http/simple.js type=bytes length=1024 c=500: ./node-original: 2008.7 ./node: 1987 ...................... 1.10%
http/simple.js type=bytes length=102400 c=50: ./node-original: 332.78 ./node: 328.77 ................... 1.22%
http/simple.js type=bytes length=102400 c=500: ./node-original: 316.08 ./node: 304.85 .................. 3.68%
http/simple.js type=buffer length=4 c=50: ./node-original: 2199.9 ./node: 2113.9 ....................... 4.07%
http/simple.js type=buffer length=4 c=500: ./node-original: 2034.8 ./node: 1999.5 ...................... 1.76%
http/simple.js type=buffer length=1024 c=50: ./node-original: 2177.1 ./node: 2168.3 .................... 0.40%
http/simple.js type=buffer length=1024 c=500: ./node-original: 2078.2 ./node: 1975.5 ................... 5.20%
http/simple.js type=buffer length=102400 c=50: ./node-original: 1265.9 ./node: 1255.6 .................. 0.82%
http/simple.js type=buffer length=102400 c=500: ./node-original: 1223.5 ./node: 1230.3 ................ -0.56%

indutny · 2013-04-14T12:33:47Z

@bnoordhuis @isaacs this coalesce thing seems to be to broad and I agree with @bnoordhuis that using it for every net write is a bad idea. I've force pushed this PR and now it implements bulk method for writable stream and consumes it in http module. Please let me know how it looks for you!

indutny · 2013-04-14T17:42:46Z

Fresh benchmark results:


http/chunked.js num=1 size=1 c=100: ./node-with-changes: 1709.2 ./node-original: 1325.6 ........................... 28.94%
http/chunked.js num=1 size=64 c=100: ./node-with-changes: 1645.6 ./node-original: 1304.3 .......................... 26.17%
http/chunked.js num=1 size=256 c=100: ./node-with-changes: 1666.8 ./node-original: 1292.9 ......................... 28.92%
http/chunked.js num=4 size=1 c=100: ./node-with-changes: 736.32 ./node-original: 452.2 ............................ 62.83%
http/chunked.js num=4 size=64 c=100: ./node-with-changes: 765.07 ./node-original: 452.06 .......................... 69.24%
http/chunked.js num=4 size=256 c=100: ./node-with-changes: 723.93 ./node-original: 448.04 ......................... 61.58%
http/chunked.js num=8 size=1 c=100: ./node-with-changes: 407.65 ./node-original: 243.07 ........................... 67.71%
http/chunked.js num=8 size=64 c=100: ./node-with-changes: 413.22 ./node-original: 240.91 .......................... 71.53%
http/chunked.js num=8 size=256 c=100: ./node-with-changes: 419.79 ./node-original: 230.4 .......................... 82.20%
http/chunked.js num=16 size=1 c=100: ./node-with-changes: 229.7 ./node-original: 122.28 ........................... 87.85%
http/chunked.js num=16 size=64 c=100: ./node-with-changes: 229.65 ./node-original: 122.25 ......................... 87.85%
http/chunked.js num=16 size=256 c=100: ./node-with-changes: 224.34 ./node-original: 118.68 ........................ 89.03%
http/cluster.js type=bytes length=4 c=50: ./node-with-changes: 3157.1 ./node-original: 3159.2 ..................... -0.06%
http/cluster.js type=bytes length=4 c=500: ./node-with-changes: 3191.3 ./node-original: 3203.4 .................... -0.38%
http/cluster.js type=bytes length=1024 c=50: ./node-with-changes: 3353.4 ./node-original: 3106.8 ................... 7.94%
http/cluster.js type=bytes length=1024 c=500: ./node-with-changes: 3076 ./node-original: 3081.9 ................... -0.19%
http/cluster.js type=bytes length=102400 c=50: ./node-with-changes: 438.89 ./node-original: 439.98 ................ -0.25%
http/cluster.js type=bytes length=102400 c=500: ./node-with-changes: 450.17 ./node-original: 463.46 ............... -2.87%
http/cluster.js type=buffer length=4 c=50: ./node-with-changes: 3375.9 ./node-original: 3244.9 ..................... 4.04%
http/cluster.js type=buffer length=4 c=500: ./node-with-changes: 3053.9 ./node-original: 3172.2 ................... -3.73%
http/cluster.js type=buffer length=1024 c=50: ./node-with-changes: 3295.4 ./node-original: 3229.4 .................. 2.05%
http/cluster.js type=buffer length=1024 c=500: ./node-with-changes: 2930.4 ./node-original: 2954.9 ................ -0.83%
http/cluster.js type=buffer length=102400 c=50: ./node-with-changes: 1708.3 ./node-original: 1672.2 ................ 2.16%
http/cluster.js type=buffer length=102400 c=500: ./node-with-changes: 1837.9 ./node-original: 1855.8 .............. -0.96%
http/end-vs-write-end.js type=asc kb=64 c=100 method=write: ./node-with-changes: 914.03 ./node-original: 939.64 ... -2.73%
http/end-vs-write-end.js type=asc kb=64 c=100 method=end  : ./node-with-changes: 1063.7 ./node-original: 1078 ..... -1.32%
http/end-vs-write-end.js type=asc kb=128 c=100 method=write: ./node-with-changes: 628.76 ./node-original: 627.02 ... 0.28%
http/end-vs-write-end.js type=asc kb=128 c=100 method=end  : ./node-with-changes: 683.45 ./node-original: 679.46 ... 0.59%
http/end-vs-write-end.js type=asc kb=256 c=100 method=write: ./node-with-changes: 386.71 ./node-original: 391.43 .. -1.21%
http/end-vs-write-end.js type=asc kb=256 c=100 method=end  : ./node-with-changes: 409.83 ./node-original: 406.41 ... 0.84%
http/end-vs-write-end.js type=asc kb=1024 c=100 method=write: ./node-with-changes: 101.64 ./node-original: 102.08 . -0.43%
http/end-vs-write-end.js type=asc kb=1024 c=100 method=end  : ./node-with-changes: 103.6 ./node-original: 99.957 ... 3.65%
http/end-vs-write-end.js type=utf kb=64 c=100 method=write: ./node-with-changes: 1067.8 ./node-original: 1085.4 ... -1.63%
http/end-vs-write-end.js type=utf kb=64 c=100 method=end  : ./node-with-changes: 1280.5 ./node-original: 1339 ..... -4.38%
http/end-vs-write-end.js type=utf kb=128 c=100 method=write: ./node-with-changes: 765.84 ./node-original: 746.59 ... 2.58%
http/end-vs-write-end.js type=utf kb=128 c=100 method=end  : ./node-with-changes: 818.96 ./node-original: 837.74 .. -2.24%
http/end-vs-write-end.js type=utf kb=256 c=100 method=write: ./node-with-changes: 492.8 ./node-original: 492.54 .... 0.05%
http/end-vs-write-end.js type=utf kb=256 c=100 method=end  : ./node-with-changes: 530.3 ./node-original: 528.3 ..... 0.38%
http/end-vs-write-end.js type=utf kb=1024 c=100 method=write: ./node-with-changes: 155.13 ./node-original: 156.87 . -1.11%
http/end-vs-write-end.js type=utf kb=1024 c=100 method=end  : ./node-with-changes: 159.63 ./node-original: 158.48 .. 0.73%
http/end-vs-write-end.js type=buf kb=64 c=100 method=write: ./node-with-changes: 1297.5 ./node-original: 1009.6 ... 28.52%
http/end-vs-write-end.js type=buf kb=64 c=100 method=end  : ./node-with-changes: 1637.3 ./node-original: 1691 ..... -3.17%
http/end-vs-write-end.js type=buf kb=128 c=100 method=write: ./node-with-changes: 1112.5 ./node-original: 861.66 .. 29.11%
http/end-vs-write-end.js type=buf kb=128 c=100 method=end  : ./node-with-changes: 1111 ./node-original: 868.84 .... 27.88%
http/end-vs-write-end.js type=buf kb=256 c=100 method=write: ./node-with-changes: 944.22 ./node-original: 738.03 .. 27.94%
http/end-vs-write-end.js type=buf kb=256 c=100 method=end  : ./node-with-changes: 930 ./node-original: 730.26 ..... 27.35%
http/end-vs-write-end.js type=buf kb=1024 c=100 method=write: ./node-with-changes: 440.38 ./node-original: 346.51 . 27.09%
http/end-vs-write-end.js type=buf kb=1024 c=100 method=end  : ./node-with-changes: 445.63 ./node-original: 349.54 . 27.49%
http/simple.js type=bytes length=4 c=50: ./node-with-changes: 2189.8 ./node-original: 2335.6 ...................... -6.24%
http/simple.js type=bytes length=4 c=500: ./node-with-changes: 2165.4 ./node-original: 2126.9 ...................... 1.81%
http/simple.js type=bytes length=1024 c=50: ./node-with-changes: 2247.2 ./node-original: 2160.6 .................... 4.01%
http/simple.js type=bytes length=1024 c=500: ./node-with-changes: 1978 ./node-original: 2059.4 .................... -3.95%
http/simple.js type=bytes length=102400 c=50: ./node-with-changes: 310.34 ./node-original: 310.36 ................. -0.01%
http/simple.js type=bytes length=102400 c=500: ./node-with-changes: 295.44 ./node-original: 297.76 ................ -0.78%
http/simple.js type=buffer length=4 c=50: ./node-with-changes: 2513.6 ./node-original: 2295.8 ...................... 9.49%
http/simple.js type=buffer length=4 c=500: ./node-with-changes: 1952.6 ./node-original: 1919.5 ..................... 1.72%
http/simple.js type=buffer length=1024 c=50: ./node-with-changes: 2333 ./node-original: 2196.6 ..................... 6.21%
http/simple.js type=buffer length=1024 c=500: ./node-with-changes: 1990.3 ./node-original: 1939 .................... 2.64%
http/simple.js type=buffer length=102400 c=50: ./node-with-changes: 1291.6 ./node-original: 1264.4 ................. 2.16%
http/simple.js type=buffer length=102400 c=500: ./node-with-changes: 1237.2 ./node-original: 1215.8 ................ 1.76%

isaacs · 2013-04-21T17:18:32Z

lib/_stream_writable.js

+        !state.bufferProcessing &&
+        state.buffer.length)
+      clearBuffer(this, state);
+  } else {


This error message is kind of weird. Also, why nextTick? Seems like it should emit right now. Maybe something like 'Cannot uncork, not corked'?

indutny · 2013-04-26T15:45:20Z

@isaacs updated, following tips from our previous discussion. Please check and let me know if I forgot anything.

isaacs · 2013-04-27T05:55:32Z

Still a few nits to pick...

writev() should get an array of {chunk,encoding} objects, rather than 2 parallel arrays. (Ie, Writable should just pass the buffer list, more or less as-is, I think.)
long variadic functions are a no-no.

I'll add a patch tomorrow to help out with this.

isaacs · 2013-04-27T06:23:39Z

Top two patches here change the API to what we decided on the mailing list: https://github.com/isaacs/node/commits/writev

It'd still be good to get rid of the variadic functions and long functions. It's better to be less dry than to have variable function signatures or functions that are too long, or return early. This is hot code, need to kiss up to V8.

indutny · 2013-04-27T12:00:42Z

@isaacs thanks, I've merged your changes into my branch. However I took slightly different approach for net.js and made writev() accept pairs of data/encoding as an argument (for consistency reasons): handle.writev([ [data, enc], [data, enc], ... ]).

Add Writev method to StreamWrap class for writing mixed array of strings and buffers. Expose this method for TCP class.

indutny mentioned this pull request Apr 8, 2013

Streams2 + net + http: writev #5246

Closed

isaacs reviewed Apr 21, 2013
View reviewed changes

streams: introduce .cork/.uncork/._writev

21ed8df

indutny added 4 commits April 27, 2013 21:03

net: implement ._writev for .cork/uncork() support

60ed2c5

Add Writev method to StreamWrap class for writing mixed array of strings and buffers. Expose this method for TCP class.

http: write buffers when chunked to embrace writev

df8a4f8

test: fix timing sensitive tests

0c72936

benchmark: add chunked-encoding benchmark

f2d5cea

isaacs merged commit f2d5cea into nodejs:master Apr 28, 2013

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Feature/writev master #5257

Feature/writev master #5257

indutny commented Apr 8, 2013

indutny commented Apr 9, 2013

indutny commented Apr 9, 2013

indutny commented Apr 9, 2013

indutny commented Apr 9, 2013

indutny commented Apr 14, 2013

indutny commented Apr 14, 2013

isaacs Apr 21, 2013

indutny commented Apr 26, 2013

isaacs commented Apr 27, 2013

isaacs commented Apr 27, 2013

indutny commented Apr 27, 2013

Feature/writev master #5257

Feature/writev master #5257

Conversation

indutny commented Apr 8, 2013

indutny commented Apr 9, 2013

indutny commented Apr 9, 2013

indutny commented Apr 9, 2013

indutny commented Apr 9, 2013

indutny commented Apr 14, 2013

indutny commented Apr 14, 2013

isaacs Apr 21, 2013

Choose a reason for hiding this comment

indutny commented Apr 26, 2013

isaacs commented Apr 27, 2013

isaacs commented Apr 27, 2013

indutny commented Apr 27, 2013