src: remove pushValueToArray and setupProcessObject #24264

joyeecheung · 2018-11-09T06:07:15Z

This PR removes

pushValueToArray
NODE_PUSH_VAL_TO_ARRAY_MAX
env->env->push_values_to_array_function()

In favor of the new V8 C++ API that constructs an Array from a C++ array directly.

Also removes setupProcessObject since by now it is only doing the push_values_to_array_function setup.
Also added a test for os.cpus() values.

Checklist

make -j4 test (UNIX), or vcbuild test (Windows) passes
tests and/or benchmarks are included
commit message follows commit guidelines

nodejs-github-bot · 2018-11-09T06:07:16Z

@joyeecheung build started: https://ci.nodejs.org/blue/organizations/jenkins/node-test-pull-request-lite-pipeline/detail/node-test-pull-request-lite-pipeline/1550/pipeline

joyeecheung · 2018-11-09T06:09:30Z

Some local benchmark results:

                               confidence improvement accuracy (*)   (**)  (***)
 process/bench-env.js n=100000         **      1.99 %       ±1.39% ±1.84% ±2.40%

                    confidence improvement accuracy (*)   (**)  (***)
 os/cpus.js n=30000                -0.13 %       ±0.77% ±1.02% ±1.33%

                                        confidence improvement accuracy (*)   (**)  (***)
 http2/headers.js nheaders=0 n=1000                    0.33 %       ±1.06% ±1.40% ±1.83%
 http2/headers.js nheaders=10 n=1000                   0.44 %       ±1.03% ±1.37% ±1.79%
 http2/headers.js nheaders=100 n=1000                 -0.25 %       ±0.97% ±1.29% ±1.68%
 http2/headers.js nheaders=1000 n=1000         **      1.96 %       ±1.40% ±1.86% ±2.42%

The bench-parser results are rather flaky, and the benchmark is written in a strange way - it executes the same parser on the same payload in a hot loop, which almost never happen in the real world. When I turn the number of loop iteration down to 100 (probably the optimizing compiler doesn't kick in or doesn't do OSR doesn't optimize as aggressively that way) the improvement seems consistent, but I am not quite sure about that either. Then HTTP2 header parser that are migrated in the same way showed no significant impact, however, but that's benchmarked differently (it's not in a hot loop, instead it parse the headers sequentially in callbacks)

Intuitively the new API is doing what has to be done either in JS or C++ - it constructs an Array out of a FixedArray with all the elements readily packed inside, so it shouldn't be slower than the old way (which calls into JS), the impact seems to come from how the new approach affect what gets optimized in a hot loop which rarely happen in real word, so I am inclined to just ignore the results in http/bench-parser, but someone who has a better knowledge on the optimizing compiler may have a better insight than I am (cc @bmeurer ?).

For reference, something like #24125 shows significant improvement when the length of the array is long enough and the call cannot be optimized (e.g. unconditionally done in C++ land without kHasNoSideEffect notation)

From `node benchmark/compare.js --new ./node_new --old ./node_old --filter parser http`

                                       confidence improvement accuracy (*)   (**)   (***)
 http/bench-parser.js n=100000 len=16        ***    -13.15 %       ±7.09% ±9.44% ±12.29%
 http/bench-parser.js n=100000 len=32          *     -9.01 %       ±6.81% ±9.07% ±11.80%
 http/bench-parser.js n=100000 len=4         ***    -22.69 %       ±7.29% ±9.70% ±12.65%
 http/bench-parser.js n=100000 len=8         ***    -17.32 %       ±6.99% ±9.31% ±12.12%

Another time from `node benchmark/compare.js --new ./node_new --old ./node_old --filter parser http`

                                      confidence improvement accuracy (*)    (**)   (***)
 http/bench-parser.js n=100000 len=16               -12.83 %      ±17.71% ±23.56% ±30.67%
 http/bench-parser.js n=100000 len=32                -9.39 %      ±17.83% ±23.73% ±30.88%
 http/bench-parser.js n=100000 len=4           *    -21.00 %      ±17.30% ±23.04% ±30.01%
 http/bench-parser.js n=100000 len=8                -16.33 %      ±17.42% ±23.18% ±30.17%

Setting n=100:

                                   confidence improvement accuracy (*)   (**)  (***)
 http/bench-parser.js n=100 len=16        ***     27.89 %       ±2.98% ±3.98% ±5.19%
 http/bench-parser.js n=100 len=32        ***     40.82 %       ±2.65% ±3.56% ±4.70%
 http/bench-parser.js n=100 len=4           *      5.47 %       ±4.81% ±6.41% ±8.37%
 http/bench-parser.js n=100 len=8         ***     17.01 %       ±3.77% ±5.07% ±6.73%

bmeurer

Not sure about the benchmark results TBH.

bmeurer · 2018-11-09T08:21:07Z

src/node_http2.cc

-  Local<Value> holder = Array::New(isolate);
-  Local<Function> fn = env()->push_values_to_array_function();
-  Local<Value> argv[NODE_PUSH_VAL_TO_ARRAY_MAX];
+  std::vector<Local<Value>> origin_v;


Since you already know the capacity needed for this, how about pre-allocating it here?

Good idea! Thanks for catching that

bmeurer · 2018-11-09T08:21:30Z

src/node_http_parser.cc

-        fn->Call(env()->context(), headers, j * 2, argv).ToLocalChecked();
-      }
-    } while (i < num_values_);
+    std::vector<Local<Value>> headers_v;


Same here, how about preallocating the appropriate capacity?

Now come to think of it, this one can actually be an array since its capacity cannot be bigger than 32*2, maybe that'll speed it up a bit more..

~~I'd trust std::vector. No need to over-optimize.~~ Are the benchmark around this? 😮

Instead of calling into JS from C++ to push values into an array, use the new Array::New API that takes a pointer and a length directly.

The usage of NODE_PUSH_VAL_TO_ARRAY_MAX and push_values_to_array_function has all been removed in favor of the new Array::New API that takes a C++ array. Remove the unused code.

joyeecheung · 2018-11-11T21:28:37Z

Apparently using a fixed-size array in the http parser does the trick, though the impact shown from bench-parser.js is still influenced by the number of iterations, and the accuracy is still not too high when n=1e5 - I don't really know what to make out of this but I guess it doesn't really hurt ¯\(ツ)/¯

                                      confidence improvement accuracy (*)   (**)  (***)
 http/bench-parser.js n=100000 len=16        ***     10.44 %       ±3.90% ±5.20% ±6.76%
 http/bench-parser.js n=100000 len=32        ***     11.36 %       ±3.78% ±5.04% ±6.59%
 http/bench-parser.js n=100000 len=4           *      6.81 %       ±5.39% ±7.17% ±9.33%
 http/bench-parser.js n=100000 len=8         ***     11.29 %       ±3.82% ±5.09% ±6.63%

when n=100 (somehow accuracy is a little bit better):

                                   confidence improvement accuracy (*)   (**)  (***)
 http/bench-parser.js n=100 len=16        ***     47.65 %       ±2.68% ±3.57% ±4.65%
 http/bench-parser.js n=100 len=32        ***     62.44 %       ±3.34% ±4.46% ±5.85%
 http/bench-parser.js n=100 len=4         ***     21.83 %       ±4.10% ±5.47% ±7.17%
 http/bench-parser.js n=100 len=8         ***     41.69 %       ±3.60% ±4.83% ±6.36%

Other benchmark results:

                               confidence improvement accuracy (*)   (**)  (***)
 process/bench-env.js n=100000         **      3.65 %       ±2.36% ±3.14% ±4.09%

                    confidence improvement accuracy (*)   (**)  (***)
 os/cpus.js n=30000                -0.27 %       ±2.07% ±2.76% ±3.59%

                                       confidence improvement accuracy (*)   (**)  (***)
 http2/headers.js nheaders=0 n=1000                   -0.42 %       ±2.30% ±3.06% ±3.99%
 http2/headers.js nheaders=10 n=1000                   0.89 %       ±2.10% ±2.80% ±3.65%
 http2/headers.js nheaders=100 n=1000                  1.00 %       ±2.16% ±2.87% ±3.74%
 http2/headers.js nheaders=1000 n=1000          *      1.61 %       ±1.55% ±2.07% ±2.69%

joyeecheung · 2018-11-11T21:33:20Z

CI: https://ci.nodejs.org/job/node-test-pull-request/18540/

refack

🎉
(just have some style nits)

lib/os.js

src/node_http2.cc

refack · 2018-11-11T20:53:53Z

src/node_http2.cc

+  size_t headers_size = headers.size();
+  std::vector<Local<Value>> headers_v(headers_size * 2);
+  for (size_t i = 0; i < headers_size; ++i) {
+    nghttp2_header item = headers[i];


BTW: this is an unneeded copy

Good catch! (though I realize it's originally that way :/)

I think the compiler should be able optimize this away anyway?

src/node_http2.cc

refack · 2018-11-11T21:11:18Z

src/node_http2.cc

  }
-
+  Local<Value> holder = Array::New(isolate, origin_v.data(), nov);


/s/nov/origin_v.size()/ for correctness and readability

src/node_http_parser.cc

src/node_os.cc

lib/os.js

…rame

lib/os.js

addaleax

Nice work!

And, to be explicit, I’m 👍 on using indexed for-loops here as well (for the reasons @joyeecheung is mentioning).

addaleax · 2018-11-12T12:32:04Z

src/node_http2.cc

+  size_t headers_size = headers.size();
+  std::vector<Local<Value>> headers_v(headers_size * 2);
+  for (size_t i = 0; i < headers_size; ++i) {
+    nghttp2_header item = headers[i];


I think the compiler should be able optimize this away anyway?

addaleax · 2018-11-12T12:33:51Z

src/node_http2.cc

-    if (j > 0)
-      fn->Call(context, holder, j, argv).ToLocalChecked();
+  for (size_t i = 0; i < nov; ++i) {
+    auto entry = origin->ov[i];


I know this is copy-pasted, but I think getting rid of this auto might make things more readable :)

joyeecheung · 2018-11-13T13:35:15Z

Interesting results: I tried using reserve and then push_back or emplace_back in CreateHeaders (the hottest path in this patch), and they both cause a slight performance regression compared to assigning and indexing into the vector with high accuracy.

push_back:

                                      confidence improvement accuracy (*)   (**)  (***)
 http/bench-parser.js n=100000 len=16         **     -2.38 %       ±1.57% ±2.10% ±2.74%
 http/bench-parser.js n=100000 len=32        ***     -3.26 %       ±1.55% ±2.07% ±2.69%
 http/bench-parser.js n=100000 len=4         ***     -6.79 %       ±1.71% ±2.28% ±2.98%
 http/bench-parser.js n=100000 len=8         ***     -3.68 %       ±2.09% ±2.78% ±3.63%

emplace_back:

                                      confidence improvement accuracy (*)   (**)  (***)
 http/bench-parser.js n=100000 len=16                -0.92 %       ±1.81% ±2.41% ±3.13%
 http/bench-parser.js n=100000 len=32        ***     -5.43 %       ±2.19% ±2.92% ±3.80%
 http/bench-parser.js n=100000 len=4         ***     -4.89 %       ±1.88% ±2.52% ±3.29%
 http/bench-parser.js n=100000 len=8         ***     -2.76 %       ±1.58% ±2.11% ±2.75%

I suspect RVO is somehow not in effect with the way v8::Local<> is implemented/templated? Note that v8::Local<>'s default constructor merely initializes a pointer to 0 and the copy constructor doesn't do much either. But I prefer indexing into the vector with self-documenting equations anyway, so I didn't look into that any further.

Fixed a few nits with const references and a few comments. CI: https://ci.nodejs.org/job/node-test-pull-request/18576/

Instead of calling into JS from C++ to push values into an array, use the new Array::New API that takes a pointer and a length directly. PR-URL: #24264 Reviewed-By: Anna Henningsen <[email protected]> Reviewed-By: Daniel Bevenius <[email protected]> Reviewed-By: Refael Ackermann <[email protected]>

The usage of NODE_PUSH_VAL_TO_ARRAY_MAX and push_values_to_array_function has all been removed in favor of the new Array::New API that takes a C++ array. Remove the unused code. PR-URL: #24264 Reviewed-By: Anna Henningsen <[email protected]> Reviewed-By: Daniel Bevenius <[email protected]> Reviewed-By: Refael Ackermann <[email protected]>

Instead of calling into JS from C++ to push values into an array, use the new Array::New API that takes a pointer and a length directly. PR-URL: #24264 Reviewed-By: Anna Henningsen <[email protected]> Reviewed-By: Daniel Bevenius <[email protected]> Reviewed-By: Refael Ackermann <[email protected]>

The usage of NODE_PUSH_VAL_TO_ARRAY_MAX and push_values_to_array_function has all been removed in favor of the new Array::New API that takes a C++ array. Remove the unused code. PR-URL: #24264 Reviewed-By: Anna Henningsen <[email protected]> Reviewed-By: Daniel Bevenius <[email protected]> Reviewed-By: Refael Ackermann <[email protected]>

Instead of calling into JS from C++ to push values into an array, use the new Array::New API that takes a pointer and a length directly. PR-URL: nodejs#24264 Reviewed-By: Anna Henningsen <[email protected]> Reviewed-By: Daniel Bevenius <[email protected]> Reviewed-By: Refael Ackermann <[email protected]>

The usage of NODE_PUSH_VAL_TO_ARRAY_MAX and push_values_to_array_function has all been removed in favor of the new Array::New API that takes a C++ array. Remove the unused code. PR-URL: nodejs#24264 Reviewed-By: Anna Henningsen <[email protected]> Reviewed-By: Daniel Bevenius <[email protected]> Reviewed-By: Refael Ackermann <[email protected]>

codebytere · 2019-01-12T21:45:02Z

@joyeecheung can/should this be backported to v10.x?

nodejs-github-bot added c++ Issues and PRs that require attention from people who are familiar with C++. lib / src Issues and PRs related to general changes in the lib or src directory. labels Nov 9, 2018

bmeurer reviewed Nov 9, 2018

View reviewed changes

joyeecheung added 6 commits November 12, 2018 03:44

http2: remove pushValueToArray in Http2Session::HandleHeadersFrame

a9f8f33

Instead of calling into JS from C++ to push values into an array, use the new Array::New API that takes a pointer and a length directly.

http2: remove pushValueToArray in Http2Session::HandleOriginFrame

3e3e9b7

Instead of calling into JS from C++ to push values into an array, use the new Array::New API that takes a pointer and a length directly.

http_parser: remove pushValueToArray in CreateHeaders

fa1c240

Instead of calling into JS from C++ to push values into an array, use the new Array::New API that takes a pointer and a length directly.

process: remove pushValueToArray in EnvEnumerator

a96a222

Instead of calling into JS from C++ to push values into an array, use the new Array::New API that takes a pointer and a length directly.

process: remove pushValueToArray in GetActiveRequests

d79ed5d

Instead of calling into JS from C++ to push values into an array, use the new Array::New API that takes a pointer and a length directly.

process: remove pushValueToArray in GetActiveHandles

70fa3c1

Instead of calling into JS from C++ to push values into an array, use the new Array::New API that takes a pointer and a length directly.

joyeecheung force-pushed the remove-push branch from 3cb8bf9 to cc511fb Compare November 11, 2018 19:44

joyeecheung added 2 commits November 12, 2018 05:14

os: do not call into JS to push values to an array in GetCPUInfo

a54d0c6

Instead of calling into JS from C++ to push values into an array, use the new Array::New API that takes a pointer and a length directly.

src: remove pushValueToArray and SetupProcessObject

db5e5fa

The usage of NODE_PUSH_VAL_TO_ARRAY_MAX and push_values_to_array_function has all been removed in favor of the new Array::New API that takes a C++ array. Remove the unused code.

joyeecheung force-pushed the remove-push branch from cc511fb to db5e5fa Compare November 11, 2018 21:21

fixup! process: remove pushValueToArray in EnvEnumerator

0b3568c

refack reviewed Nov 11, 2018

View reviewed changes

lib/os.js Show resolved Hide resolved

fixup! http2: remove pushValueToArray in Http2Session::HandleHeadersF…

a44e978

…rame

refack reviewed Nov 11, 2018

View reviewed changes

lib/os.js Show resolved Hide resolved

This comment has been minimized.

Sign in to view

addaleax approved these changes Nov 12, 2018

View reviewed changes

danbev approved these changes Nov 13, 2018

View reviewed changes

joyeecheung mentioned this pull request Nov 13, 2018

doc: make C++ core guidelines a reference #24315

Closed

joyeecheung added 2 commits November 13, 2018 21:01

fixup! http2: remove pushValueToArray in Http2Session::HandleOriginFrame

fcaf05a

fixup! os: do not call into JS to push values to an array in GetCPUInfo

07b5417

joyeecheung added the author ready PRs that have at least one approval, no pending requests for changes, and a CI started. label Nov 13, 2018

BridgeAR mentioned this pull request Nov 14, 2018

Release proposal: v11.2 #24350

Merged

This was referenced Nov 17, 2018

doc: include undefined in os.cpus() return values #24408

Closed

os needs fix to detect CPUs or fall back to 1 #19022

Closed

targos mentioned this pull request Nov 28, 2018

src: fix compiler warnings #23880

Closed

2 tasks

codebytere added the backport-requested-v10.x label Jan 12, 2019

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

src: remove pushValueToArray and setupProcessObject #24264

src: remove pushValueToArray and setupProcessObject #24264

joyeecheung commented Nov 9, 2018

nodejs-github-bot commented Nov 9, 2018

joyeecheung commented Nov 9, 2018 •

edited

Loading

bmeurer left a comment

bmeurer Nov 9, 2018

joyeecheung Nov 11, 2018

bmeurer Nov 9, 2018

joyeecheung Nov 11, 2018

refack Nov 11, 2018 •

edited

Loading

joyeecheung commented Nov 11, 2018

joyeecheung commented Nov 11, 2018

refack left a comment

refack Nov 11, 2018

joyeecheung Nov 11, 2018

addaleax Nov 12, 2018

refack Nov 11, 2018

This comment has been minimized.

addaleax left a comment

addaleax Nov 12, 2018

addaleax Nov 12, 2018

joyeecheung commented Nov 13, 2018 •

edited

Loading

codebytere commented Jan 12, 2019

		}

		Local<Value> holder = Array::New(isolate, origin_v.data(), nov);

src: remove pushValueToArray and setupProcessObject #24264

src: remove pushValueToArray and setupProcessObject #24264

Conversation

joyeecheung commented Nov 9, 2018

Checklist

nodejs-github-bot commented Nov 9, 2018

joyeecheung commented Nov 9, 2018 • edited Loading

bmeurer 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

refack Nov 11, 2018 • edited Loading

Choose a reason for hiding this comment

joyeecheung commented Nov 11, 2018

joyeecheung commented Nov 11, 2018

refack 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

This comment has been minimized.

addaleax left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

joyeecheung commented Nov 13, 2018 • edited Loading

codebytere commented Jan 12, 2019

joyeecheung commented Nov 9, 2018 •

edited

Loading

refack Nov 11, 2018 •

edited

Loading

joyeecheung commented Nov 13, 2018 •

edited

Loading