2010-12-19

Intel Atomの上でのJavaScriptパフォーマンス

このデータはSunSpiderAtom Z560 (VAIO P) + Ubuntu 10.10上でのFirefox 4.0b9pre (2010-12-18)とGoogle Chrome 10.0.612.1 devで走らせた結果なんだけど、すごく興味深い。FROMはChromeでTOはFirefoxね。

TEST                   COMPARISON            FROM                 TO             DETAILS

=============================================================================

** TOTAL **:           1.38x as fast     1361.8ms +/- 5.4%   987.8ms +/- 1.0%     significant

=============================================================================

  3d:                  1.41x as fast      224.0ms +/- 11.6%   158.5ms +/- 0.9%     significant
    cube:              ??                  61.7ms +/- 23.3%    64.0ms +/- 1.8%     not conclusive: might be *1.037x as slow*
    morph:             1.48x as fast       47.2ms +/- 8.7%    31.9ms +/- 3.1%     significant
    raytrace:          1.84x as fast      115.1ms +/- 9.3%    62.6ms +/- 1.1%     significant

  access:              -                  130.6ms +/- 12.0%   128.7ms +/- 2.5% 
    binary-trees:      *2.02x as slow*      9.7ms +/- 42.8%    19.6ms +/- 12.8%     significant
    fannkuch:          1.42x as fast       53.3ms +/- 18.4%    37.6ms +/- 2.7%     significant
    nbody:             1.73x as fast       48.4ms +/- 11.9%    27.9ms +/- 2.2%     significant
    nsieve:            *2.27x as slow*     19.2ms +/- 7.6%    43.6ms +/- 2.5%     significant

  bitops:              1.77x as fast       87.6ms +/- 19.7%    49.5ms +/- 4.9%     significant
    3bit-bits-in-byte: 2.79x as fast        7.8ms +/- 5.8%     2.8ms +/- 10.8%     significant
    bits-in-byte:      ??                  17.8ms +/- 37.5%    19.6ms +/- 3.1%     not conclusive: might be *1.101x as slow*
    bitwise-and:       4.19x as fast       30.6ms +/- 22.5%     7.3ms +/- 6.6%     significant
    nsieve-bits:       1.59x as fast       31.4ms +/- 14.6%    19.8ms +/- 8.3%     significant

  controlflow:         1.49x as fast       12.7ms +/- 14.0%     8.5ms +/- 5.9%     significant
    recursive:         1.49x as fast       12.7ms +/- 14.0%     8.5ms +/- 5.9%     significant

  crypto:              1.65x as fast      111.6ms +/- 6.8%    67.6ms +/- 1.1%     significant
    aes:               1.61x as fast       59.7ms +/- 14.7%    37.1ms +/- 2.8%     significant
    md5:               1.64x as fast       26.0ms +/- 7.3%    15.9ms +/- 1.4%     significant
    sha1:              1.77x as fast       25.9ms +/- 7.1%    14.6ms +/- 3.4%     significant

  date:                1.50x as fast      241.3ms +/- 3.7%   161.3ms +/- 1.4%     significant
    format-tofte:      *1.116x as slow*    84.7ms +/- 8.2%    94.5ms +/- 1.2%     significant
    format-xparb:      2.34x as fast      156.6ms +/- 4.6%    66.8ms +/- 4.2%     significant

  math:                ??                  88.1ms +/- 4.4%    91.6ms +/- 1.3%     not conclusive: might be *1.040x as slow*
    cordic:            *1.47x as slow*     19.4ms +/- 10.7%    28.5ms +/- 2.7%     significant
    partial-sums:      1.27x as fast       51.0ms +/- 5.3%    40.1ms +/- 1.0%     significant
    spectral-norm:     *1.30x as slow*     17.7ms +/- 6.0%    23.0ms +/- 2.1%     significant

  regexp:              1.34x as fast       50.2ms +/- 8.4%    37.5ms +/- 7.6%     significant
    dna:               1.34x as fast       50.2ms +/- 8.4%    37.5ms +/- 7.6%     significant

  string:              1.46x as fast      415.7ms +/- 12.6%   284.6ms +/- 1.9%     significant
    base64:            1.44x as fast       30.2ms +/- 6.4%    21.0ms +/- 6.2%     significant
    fasta:             1.25x as fast       60.7ms +/- 5.1%    48.5ms +/- 2.6%     significant
    tagcloud:          1.31x as fast      114.1ms +/- 9.5%    87.2ms +/- 1.4%     significant
    unpack-code:       1.60x as fast      137.3ms +/- 22.0%    85.6ms +/- 2.5%     significant
    validate-input:    1.74x as fast       73.4ms +/- 20.7%    42.3ms +/- 6.5%     significant

いろんなサイトで見るデータってのは、Core 2 DuoとかCore i7で走らせたものだから、互角 (新しいV8を採用してるChromeだからV8の方が速いかも?)なんだけど、Chrome OSがターゲットにしてるであろうIntel Atomプロセッサだと、V8って遅いんだよね。実際の理由は調べてないけど、CPUキャッシュの量が少ないからとかなのかもしれない。

2010-12-03

最近の作ったプレゼン資料

これは、KOF 2010で使った物で、Fennecの話。

インストールサイズを少なくするために、libdlさえもカスタム版になってしまってるとか、NDKをそのまま使えない(使い物にならないともいう)から、独自版になっていってるという話とか、合計60分話せる資料を20分でやる暴挙でしたが。

これは、Firefox Developers Conference 2010の資料。上はFirefox 4の話。下は、LTで使ったモバイルの話。