Bug 79184 – Implement SpinLockMutexLocker on WTF and optimize some parts of JavaScriptCore using it

• Component: Web Template Framework
• Status: UNCONFIRMED

[Kwonjin Jeong]

Implement SpinLockLocker on WTF and optimize some parts of JavaScriptCore using SpinLockLocker

[Kwonjin Jeong]

Change the class name to SpinLockMutexLocker.

[Kwonjin Jeong]

Created attachment 128114 [details]
Patch

[Alexey Proskuryakov]

How did you confirm that this patch is an optimization?

[Kwonjin Jeong]

(In reply to comment #3)
> How did you confirm that this patch is an optimization?

At first, I compare the performance just between MutexLocker and SpinLockMutexLocker. Upon examination, SpinLockMutexLocker is two times faster than MutexLocker.
I created a performance test program that has a loop within a simple calculation and the Lockers. Then, the program loops one billion times.

And then, I applied SpinLockMutexLocker on some parts of JavaScriptCore, then ran 'run-sunspider', and then compare the results.
Actually, the result using SpinLockMutexLocker is 0.5% ~ 1.0% faster than MutexLocker's - it's within the margin of error.
But I guess that the performance can be better by using SpinLockMutexLocker on some more parts.


My test environment and test result is as written below.
Port: QtWebKit
OS: Linux (Ubuntu 11.10)
CPU: Intel(R) Xeon(R) CPU X5650 @ 2.67GHz (6 core)
RAM: 12GiB

Loop test result:
Mutex        : 17047.357910 ms
SpinLockMutex: 7897.659180 ms
Simple Loop  : 656.346680 ms
Sum          : 565039360

Sunspider result:



The test program is WebKit/Source/JavaScriptCore/jsc.cpp with some modification. I just modify jscmain() function.

int jscmain(int argc, char** argv, JSGlobalData* globalData)
{
    int i;
    int sum = 0;
    Mutex mutex;
    SpinLockMutex spinLockMutex;
    double startTime;
    double endTimeMutexLocker;
    double endTimeSpinLock;
    double endTimeSimpleLoop;
    startTime = currentTime();
    for (i = 0; i < 1000000000; i++) {
        MutexLocker locker(mutex);
        sum += i;
    }
    endTimeMutexLocker = currentTime();
    for (i = 0; i < 1000000000; i++) {
        SpinLockMutexLocker locker(spinLockMutex);
        sum += i;
    }
    endTimeSpinLock = currentTime();
    for (i = 0; i < 1000000000; i++) {
        sum += i;
    }
    endTimeSimpleLoop = currentTime();
    printf("Mutex        : %lf ms\n", endTimeMutexLocker * 1000.0 - startTime * 1000.0);
    printf("SpinLockMutex: %lf ms\n", endTimeSpinLock * 1000.0 - endTimeMutexLocker * 1000.0);
    printf("Simple Loop  : %lf ms\n", endTimeSimpleLoop * 1000.0 - endTimeSpinLock * 1000.0);
    printf("Sum          : %d\n", sum);
    return 0;
}

[Kwonjin Jeong]

(In reply to comment #3)
> How did you confirm that this patch is an optimization?

I attach the sunspider results below.

MutexLocker
============================================
RESULTS (means and 95% confidence intervals)
--------------------------------------------
Total:                 188.7ms +/- 0.5%
--------------------------------------------

  3d:                   26.9ms +/- 1.5%
    cube:                8.0ms +/- 0.0%
    morph:              10.1ms +/- 2.2%
    raytrace:            8.8ms +/- 3.4%

  access:               17.7ms +/- 2.0%
    binary-trees:        2.0ms +/- 0.0%
    fannkuch:            8.3ms +/- 4.2%
    nbody:               4.0ms +/- 0.0%
    nsieve:              3.4ms +/- 10.9%

  bitops:               15.8ms +/- 1.9%
    3bit-bits-in-byte:   1.0ms +/- 0.0%
    bits-in-byte:        5.8ms +/- 5.2%
    bitwise-and:         3.0ms +/- 0.0%
    nsieve-bits:         6.0ms +/- 0.0%

  controlflow:           2.0ms +/- 0.0%
    recursive:           2.0ms +/- 0.0%

  crypto:               12.2ms +/- 2.5%
    aes:                 8.1ms +/- 2.8%
    md5:                 2.0ms +/- 0.0%
    sha1:                2.1ms +/- 10.8%

  date:                 22.8ms +/- 2.0%
    format-tofte:       12.2ms +/- 2.5%
    format-xparb:       10.6ms +/- 3.5%

  math:                 23.7ms +/- 2.5%
    cordic:              8.0ms +/- 0.0%
    partial-sums:       13.3ms +/- 2.6%
    spectral-norm:       2.4ms +/- 15.4%

  regexp:               11.0ms +/- 0.0%
    dna:                11.0ms +/- 0.0%

  string:               56.6ms +/- 1.4%
    base64:              4.6ms +/- 8.0%
    fasta:               8.1ms +/- 2.8%
    tagcloud:           14.1ms +/- 1.6%
    unpack-code:        23.8ms +/- 1.9%
    validate-input:      6.0ms +/- 0.0%


SpinLockMutexLocker
============================================
RESULTS (means and 95% confidence intervals)
--------------------------------------------
Total:                 187.4ms +/- 0.4%
--------------------------------------------

  3d:                   27.0ms +/- 1.2%
    cube:                8.0ms +/- 0.0%
    morph:              10.1ms +/- 2.2%
    raytrace:            8.9ms +/- 2.5%

  access:               17.5ms +/- 2.2%
    binary-trees:        2.0ms +/- 0.0%
    fannkuch:            8.1ms +/- 2.8%
    nbody:               4.0ms +/- 0.0%
    nsieve:              3.4ms +/- 10.9%

  bitops:               16.0ms +/- 0.0%
    3bit-bits-in-byte:   1.0ms +/- 0.0%
    bits-in-byte:        6.0ms +/- 0.0%
    bitwise-and:         3.0ms +/- 0.0%
    nsieve-bits:         6.0ms +/- 0.0%

  controlflow:           2.1ms +/- 10.8%
    recursive:           2.1ms +/- 10.8%

  crypto:               12.0ms +/- 0.0%
    aes:                 8.0ms +/- 0.0%
    md5:                 2.0ms +/- 0.0%
    sha1:                2.0ms +/- 0.0%

  date:                 22.3ms +/- 1.5%
    format-tofte:       12.2ms +/- 2.5%
    format-xparb:       10.1ms +/- 2.2%

  math:                 23.1ms +/- 1.0%
    cordic:              8.0ms +/- 0.0%
    partial-sums:       13.0ms +/- 0.0%
    spectral-norm:       2.1ms +/- 10.8%

  regexp:               11.0ms +/- 0.0%
    dna:                11.0ms +/- 0.0%

  string:               56.4ms +/- 0.9%
    base64:              4.4ms +/- 8.4%
    fasta:               8.0ms +/- 0.0%
    tagcloud:           14.0ms +/- 0.0%
    unpack-code:        23.9ms +/- 0.9%
    validate-input:      6.1ms +/- 3.7%

[Kwonjin Jeong]

I'm trying to fix build errors on mac and win port.

It's caused by including TCSpinLock.h but the header file isn't in PrivateHeaders.

[Alexey Proskuryakov]

Please use SunSpider built-in feature to compare the results - it will tell you if the change is statistically significant.

If the new locker is only two times as fast in a tight loop, it might not be worth adding complexity and bypassing all platform specific optimizations OS vendors add to regular mutex.

[Kwonjin Jeong]

(In reply to comment #7)
> Please use SunSpider built-in feature to compare the results - it will tell you if the change is statistically significant.
> 
> If the new locker is only two times as fast in a tight loop, it might not be worth adding complexity and bypassing all platform specific optimizations OS vendors add to regular mutex.

Thank you for your comment.

I'll try to use the feature and write the result here.

[Eric Seidel (no email)]

Comment on attachment 128114 [details]
Patch

Waiting for the sunspider-compare-results results.