string_decoder: optimize write() #1209

mscdex · 2015-03-19T21:35:57Z

By limiting property getting/setting to only where they are
absolutely necessary, we can achieve greater performance
especially with small utf8 inputs and any size base64 inputs.

ascii/binary are unaffected.

Example results for utf8 and base64 from the included benchmark:

misc/string-decode.js encoding=utf8 inlen=32 chunk=16 n=250000: ./iojs: 1946800 ./iojs-io.js-4655521: 1654400 ........... 17.67%
misc/string-decode.js encoding=utf8 inlen=32 chunk=64 n=250000: ./iojs: 2838500 ./iojs-io.js-4655521: 2545000 ........... 11.53%
misc/string-decode.js encoding=utf8 inlen=32 chunk=256 n=250000: ./iojs: 2822200 ./iojs-io.js-4655521: 2522400 .......... 11.88%
misc/string-decode.js encoding=utf8 inlen=32 chunk=1024 n=250000: ./iojs: 2829100 ./iojs-io.js-4655521: 2507800 ......... 12.81%
misc/string-decode.js encoding=utf8 inlen=128 chunk=16 n=250000: ./iojs: 508720 ./iojs-io.js-4655521: 499830 ............. 1.78%
misc/string-decode.js encoding=utf8 inlen=128 chunk=64 n=250000: ./iojs: 1016400 ./iojs-io.js-4655521: 931340 ............ 9.13%
misc/string-decode.js encoding=utf8 inlen=128 chunk=256 n=250000: ./iojs: 1250200 ./iojs-io.js-4655521: 1179200 .......... 6.02%
misc/string-decode.js encoding=utf8 inlen=128 chunk=1024 n=250000: ./iojs: 1258600 ./iojs-io.js-4655521: 1193400 ......... 5.46%
misc/string-decode.js encoding=utf8 inlen=1024 chunk=16 n=250000: ./iojs: 64944 ./iojs-io.js-4655521: 64852 .............. 0.14%
misc/string-decode.js encoding=utf8 inlen=1024 chunk=64 n=250000: ./iojs: 134380 ./iojs-io.js-4655521: 133250 ............ 0.85%
misc/string-decode.js encoding=utf8 inlen=1024 chunk=256 n=250000: ./iojs: 178810 ./iojs-io.js-4655521: 176190 ........... 1.49%
misc/string-decode.js encoding=utf8 inlen=1024 chunk=1024 n=250000: ./iojs: 215380 ./iojs-io.js-4655521: 212140 .......... 1.52%
misc/string-decode.js encoding=base64-utf8 inlen=32 chunk=16 n=250000: ./iojs: 407580 ./iojs-io.js-4655521: 307650 ...... 32.48%
misc/string-decode.js encoding=base64-utf8 inlen=32 chunk=64 n=250000: ./iojs: 1422000 ./iojs-io.js-4655521: 1120100 .... 26.96%
misc/string-decode.js encoding=base64-utf8 inlen=32 chunk=256 n=250000: ./iojs: 1416700 ./iojs-io.js-4655521: 1160400 ... 22.09%
misc/string-decode.js encoding=base64-utf8 inlen=32 chunk=1024 n=250000: ./iojs: 1339300 ./iojs-io.js-4655521: 1140600 .. 17.42%
misc/string-decode.js encoding=base64-utf8 inlen=128 chunk=16 n=250000: ./iojs: 93840 ./iojs-io.js-4655521: 84048 ....... 11.65%
misc/string-decode.js encoding=base64-utf8 inlen=128 chunk=64 n=250000: ./iojs: 383150 ./iojs-io.js-4655521: 308310 ..... 24.28%
misc/string-decode.js encoding=base64-utf8 inlen=128 chunk=256 n=250000: ./iojs: 1067600 ./iojs-io.js-4655521: 886690 ... 20.41%
misc/string-decode.js encoding=base64-utf8 inlen=128 chunk=1024 n=250000: ./iojs: 1011700 ./iojs-io.js-4655521: 926400 ... 9.21%
misc/string-decode.js encoding=base64-utf8 inlen=1024 chunk=16 n=250000: ./iojs: 13682 ./iojs-io.js-4655521: 10591 ...... 29.19%
misc/string-decode.js encoding=base64-utf8 inlen=1024 chunk=64 n=250000: ./iojs: 48976 ./iojs-io.js-4655521: 39895 ...... 22.76%
misc/string-decode.js encoding=base64-utf8 inlen=1024 chunk=256 n=250000: ./iojs: 141090 ./iojs-io.js-4655521: 120230 ... 17.35%
misc/string-decode.js encoding=base64-utf8 inlen=1024 chunk=1024 n=250000: ./iojs: 269880 ./iojs-io.js-4655521: 244070 .. 10.57%
misc/string-decode.js encoding=base64-ascii inlen=32 chunk=16 n=250000: ./iojs: 497550 ./iojs-io.js-4655521: 396600 ..... 25.45%
misc/string-decode.js encoding=base64-ascii inlen=32 chunk=64 n=250000: ./iojs: 1521700 ./iojs-io.js-4655521: 1174800 ... 29.52%
misc/string-decode.js encoding=base64-ascii inlen=32 chunk=256 n=250000: ./iojs: 1487500 ./iojs-io.js-4655521: 1174100 .. 26.69%
misc/string-decode.js encoding=base64-ascii inlen=32 chunk=1024 n=250000: ./iojs: 1401600 ./iojs-io.js-4655521: 1155800 . 21.27%
misc/string-decode.js encoding=base64-ascii inlen=128 chunk=16 n=250000: ./iojs: 142030 ./iojs-io.js-4655521: 114770 .... 23.75%
misc/string-decode.js encoding=base64-ascii inlen=128 chunk=64 n=250000: ./iojs: 499170 ./iojs-io.js-4655521: 396290 .... 25.96%
misc/string-decode.js encoding=base64-ascii inlen=128 chunk=256 n=250000: ./iojs: 1145200 ./iojs-io.js-4655521: 999490 .. 14.57%
misc/string-decode.js encoding=base64-ascii inlen=128 chunk=1024 n=250000: ./iojs: 1207200 ./iojs-io.js-4655521: 983240 . 22.77%
misc/string-decode.js encoding=base64-ascii inlen=1024 chunk=16 n=250000: ./iojs: 18163 ./iojs-io.js-4655521: 14729 ..... 23.31%
misc/string-decode.js encoding=base64-ascii inlen=1024 chunk=64 n=250000: ./iojs: 69888 ./iojs-io.js-4655521: 60726 ..... 15.09%
misc/string-decode.js encoding=base64-ascii inlen=1024 chunk=256 n=250000: ./iojs: 194780 ./iojs-io.js-4655521: 141810 .. 37.36%
misc/string-decode.js encoding=base64-ascii inlen=1024 chunk=1024 n=250000: ./iojs: 356430 ./iojs-io.js-4655521: 342270 .. 4.14%

By limiting property getting/setting to only where they are absolutely necessary, we can achieve greater performance especially with small utf8 inputs and any size base64 inputs.

rvagg · 2015-03-19T22:53:18Z

nice, I like those utf8 numbers, any idea how much property getting is impacting this?

LGTM

mscdex · 2015-03-19T23:09:37Z

I made these changes awhile ago so I can't recall for sure, but this patch is almost entirely made up of changes to avoid property lookups, so I would say that it is impacting things the most (especially in the loop in write()) versus some of the other minor non-property-related changes.

mscdex · 2015-03-20T15:30:39Z

/cc @iojs/collaborators

micnic · 2015-03-20T15:46:21Z

LGTM

petkaantonov · 2015-03-20T15:47:15Z

This should be implemented in C, we will need utf8 decoder in the future for buffer.toString() which is currently delegating to V8's utf8 decoder.

mscdex · 2015-03-20T17:25:26Z

@petkaantonov Why? Is V8 dropping their UTF-8 decoder soon?

petkaantonov · 2015-03-20T17:31:20Z

@mscdex No but instead of using NewFromUtf8 (which v8 internally decodes and copies to jsheap), we would decode it ourselves into an external buffer which doesn't take space in jsheap and also allows string sizes bigger than 290mb or so. See #649

chrisdickinson · 2015-03-20T20:03:35Z

lib/string_decoder.js

  // determine how many bytes we have to check at the end of this buffer
-  var i = (buffer.length >= 3) ? 3 : buffer.length;
+  var i = (buflen >= 3) ? 3 : buflen;


I wonder if V8 will unroll the subsequent loop (or if there's value in doing so ourselves.)

You mean for the i = 3 case? I didn't try that since I figured people would frown about code duplication between the i = 3 and i = buflen cases. I also did not see what kind performance difference it would make though.

chrisdickinson · 2015-03-20T20:10:20Z

This LGTM.

This should be implemented in C, we will need utf8 decoder in the future for buffer.toString() which is currently delegating to V8's utf8 decoder.

We'll still have to support this module if we move to a C decoder, so making it faster is probably worthwhile.

By limiting property getting/setting to only where they are absolutely necessary, we can achieve greater performance especially with small utf8 inputs and any size base64 inputs. PR-URL: #1209 Reviewed-By: Rod Vagg <[email protected]> Reviewed-By: Nicu Micleușanu <[email protected]> Reviewed-By: Chris Dickinson <[email protected]>

mscdex · 2015-03-25T04:39:01Z

Landed in 8a94581.

string_decoder: optimize write()

6d815b7

By limiting property getting/setting to only where they are absolutely necessary, we can achieve greater performance especially with small utf8 inputs and any size base64 inputs.

chrisdickinson reviewed Mar 20, 2015
View reviewed changes

mscdex added the string_decoder Issues and PRs related to the string_decoder subsystem. label Mar 25, 2015

mscdex closed this Mar 25, 2015

mscdex deleted the perf-stringdecoder branch March 25, 2015 04:39

rvagg mentioned this pull request Mar 28, 2015

Release proposal: v1.6.3 #1252

Closed

Fishrock123 mentioned this pull request Apr 9, 2015

test: windows ci timeouts related to child_process #1383

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

string_decoder: optimize write() #1209

string_decoder: optimize write() #1209

mscdex commented Mar 19, 2015

rvagg commented Mar 19, 2015

mscdex commented Mar 19, 2015

mscdex commented Mar 20, 2015

micnic commented Mar 20, 2015

petkaantonov commented Mar 20, 2015

mscdex commented Mar 20, 2015

petkaantonov commented Mar 20, 2015

chrisdickinson Mar 20, 2015

mscdex Mar 20, 2015

chrisdickinson commented Mar 20, 2015

mscdex commented Mar 25, 2015

string_decoder: optimize write() #1209

string_decoder: optimize write() #1209

Conversation

mscdex commented Mar 19, 2015

rvagg commented Mar 19, 2015

mscdex commented Mar 19, 2015

mscdex commented Mar 20, 2015

micnic commented Mar 20, 2015

petkaantonov commented Mar 20, 2015

mscdex commented Mar 20, 2015

petkaantonov commented Mar 20, 2015

chrisdickinson Mar 20, 2015

Choose a reason for hiding this comment

mscdex Mar 20, 2015

Choose a reason for hiding this comment

chrisdickinson commented Mar 20, 2015

mscdex commented Mar 25, 2015