969391 - Firefox slower than Chrome on BitcoinBench

Status: RESOLVED WORKSFORME

(Core :: JavaScript Engine: JIT, defect)

Description

[Luke Wagner [:luke]]

Patrick pointed me at this benchmark:
  https://www.bitgo.com/bitcoinbenchmark/
Nightly gets a score of 105, Chrome Dev gets 148.  As discussed in this blog post:
  http://www.belshe.com/2014/01/31/introducing-the-js-bitcoin-benchmark/
the authors expect this benchmark is representative of bitcoin computations.  (As an aside: I wonder if sites have started to embed Web Workers that mine bitcoins as an alternative to ads.)

Comment 1

[Boris Zbarsky [:bzbarsky]]

So I profiled just the first two subtests: "Address Creation" and "Sign And Verify".  The main thing jumping out at me from the profile is that 25% is under js::SetObjectElement, of which half is under JSObject::growElements.

Another 50% is in jitcode somewhere.

Another 9% is compiling jitcode, looks like.

Everything else is fairly long-tail....

Comment 2

[Jan de Mooij [:jandem]]

The website uses the Octane harness, I'll try to create a shell benchmark.

Comment 3

[Jan de Mooij [:jandem]]

Attachment: Shell version

Was pretty easy to run in the shell. I also ran the scripts through jsbeautifier.org

Comment 4

[Jan de Mooij [:jandem]]

(In reply to Boris Zbarsky [:bz] from comment #1)
> So I profiled just the first two subtests: "Address Creation" and "Sign And
> Verify".  The main thing jumping out at me from the profile is that 25% is
> under js::SetObjectElement, of which half is under JSObject::growElements.

Most of the SetObjectElement calls in "Address Creation" are for SETELEMs like this one:

    for (c = this.t - 1; c >= 0; --c)
        b[c + g + 1] = this[c] >> e | h, h = (this[c] & f) << d;

b has no elements, so the first time we run this loop LSetElementHole's OOL path will call the slow/generic SetObjectElement. Instead, the OOL path could call an helper that just does growElements, if we know there are no indexed elements on the proto. With GGC, we could even realloc the elements in JIT code.

Comment 5

[Luke Wagner [:luke]]

(In reply to Luke Wagner [:luke] from comment #0)
> (As an aside: I wonder if sites have started to embed Web Workers that mine
> bitcoins as an alternative to ads.)

Turns out, yes: http://tidbit.co.in/
And they're using asm.js to boot.  I wonder how long until we'll need to add CPU quota management functionality to FF.

Comment 6

[Guilherme Lima]

My results on Windows 7

Nightly

Address Creation 79.5
Sign And Verify 123*
SHA 89.2
Scrypt 131
AddressChain 91.2
AES 73.3

Chrome 34

Address Creation 40.8
Sign And Verify 76.2
SHA 146
Scrypt 107
AddressChain 62.1
AES 111

*made Nightly hang - on about:telemetry - Browser Hangs there is:
Hang Report #1 (11 seconds)
0xffffffff
0xffffffff

Comment 7

[Chris Peterson [:cpeterson]]

Firefox 31 handily beats Chrome 35.

Firefox 29 = 176
Beta 30    = 171
Aurora 31  = 264*
Nightly 32 = 262*
Chrome 35  = 184

Nightly 32:
    Address Creation 170
    Sign And Verify 279
    SHA 411
    Scrypt 266
    AddressChain 221
    AES 231

Chrome 35:
    Address Creation 157
    Sign And Verify 232
    SHA 243
    Scrypt 210
    AddressChain 165
    AES 199

Comment 8

[Patrick McManus [:mcmanus]]

awesome

Comment 9

[Liz Henry (:lizzard) (relman/hg->git project)]

jmaher, is this something you could use as a performance benchmark for Talos?

Comment 10

[Joel Maher ( :jmaher ) (UTC -8)]

Thanks Lizz for pointing this out.  I have filed bug 1040054 to investigate this.  We are always working on our benchmarks and tests to make them more reliable and relevant.