• Show log

    Commit

  • Hash : 41adabc4
    Author : Roman Lavrov
    Date : 2022-04-29T15:04:27

    Use app data dir instead of /sdcard for writing qpa file
    
    Appears to drastically reduce runtime variability. Example:
    
    Max shard runtime (28m 15s) + overhead (10s): 28m 24s (shard #8)
    Min shard runtime (3m 5s) + overhead (20s): 3m 24s (shard #6)
    
    -->
    
    Max shard runtime (4m 6s) + overhead (11s): 4m 17s (shard #0)
    Min shard runtime (3m 7s) + overhead (10s): 3m 17s (shard #6)
    
    --deqp-log-flush is enabled by default (only disabled for multi-process
    execution http://anglebug.com/5157) and together with
    --deqp-log-shader-sources which is also enabled by default results in a
    huge number of line-by-line flushes degradng filesystem performance.
    Example strace capture in a degraded state:
    
        [pid 21208] 10:55:51.194033 write(77, "</ShaderSource>\n   <InfoLog>", 28) = 28 <0.181028>
        [pid 21208] 10:55:51.375879 write(77, "</InfoLog>\n</FragmentShader>\n</S"..., 78) = 78 <0.179196>
        [pid 21208] 10:55:51.555790 write(77, "\" Description=\"Vertex shader com"..., 41) = 41 <0.177602>
    
    ~180ms *per line written* !!! (28, 78, 41 characters...)
    
    Under normal conditions these are way down in the microsecond range:
    
        [pid 11412] 10:55:56.689894 write(72, "</ShaderSource>\n   <InfoLog>", 28) = 28 <0.000020>
        [pid 11412] 10:55:56.689957 write(72, "</InfoLog>\n</FragmentShader>\n</S"..., 78) = 78 <0.000026>
    
    I think that the reason for this is some peculiarity of the /sdcard
    filesystem which uses fuse:
    
    % df -h /sdcard/chromium_tests_root/
    /dev/fuse        50G  3.7G   46G   8% /mnt/user/0/emulated
    
    As opposed to block/dm-N of the (non-app accessible) temp directory:
    
    % df -h /data/local/tmp/
    /dev/block/dm-9  50G  3.7G   46G   8% /data_mirror/cur_profiles
    
    And the app data directory appears to be using the same filesystem:
    
    % df -h /data/data/com.android.angle.test/
    /dev/block/dm-9  50G  3.7G   46G   8% /data_mirror/cur_profiles
    
    As far as I can tell the degradation of performance does not happen on
    this filesystem despite the huge amount of small writes. Disabling deqp
    log flushes when running on bots probably makes sense too but I'd like
    to first confirm the impact of this CL separately.
    
    Bug: chromium:1217662
    Bug: angleproject:7242
    Change-Id: I507e4c330fd4e1f4ce05f9768506f905e142f835
    Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/3615081
    Reviewed-by: Jamie Madill <jmadill@chromium.org>
    Commit-Queue: Roman Lavrov <romanl@google.com>
    

  • Properties

  • Git HTTP https://git.kmx.io/kc3-lang/angle.git
    Git SSH git@git.kmx.io:kc3-lang/angle.git
    Public access ? public
    Description

    A conformant OpenGL ES implementation for Windows, Mac, Linux, iOS and Android.

    Homepage

    Github

    Users
    thodg_m kc3_lang_org thodg_w www_kmx_io thodg thodg_l
    Tags