Changes between Version 3 and Version 4 of GpuSched


Ignore:
Timestamp:
03/18/09 10:30:18 (9 years ago)
Author:
davea
Comment:

--

Legend:

Unmodified
Added
Removed
Modified
  • GpuSched

    v3 v4  
    5656That's OK - they just run slower than if running alone.
    5757
    58 == Unresolved issues ==
    59 
    60 Apps that use GPUs use the CPU as well.
    61 The CPU part typically is a polling loop:
    62 it starts a "kernel" on the GPU,
    63 waits for it to finish (checking once per .01 sec, say)
    64 then starts another kernel.
    65 
    66 If there's a delay between when the kernel finishes
    67 and when the CPU starts another one,
    68 the GPU sits idle and the entire program runs slowly.
    69 
    70 The CPU scheduler on Windows doesn't work well,
    71 and when the CPU is overcommitted the CPU part of GPU applications
    72 doesn't run as often as it needs to in order to keep the GPU "fed".
    73 As a result the GPU is underutilized and the program runs slowly.
    74 (This seems to happen even if the GPU app is run at high priority
    75 while other apps run at low priority).
    76 
    77 If we can't resolve this we'll have to change the scheduling policy
    78 to avoid overcommitting the CPU in the presence of GPU apps.
    7958
    8059== Feature requests ==
     
    8261 * A preference to limit # of GPUs used.
    8362 * (More specific) a preference to say which specific GPU(s) can be used
    84  * Scheduler: there are a couple of places where we multiply something (e.g., max # of in-progress jobs) by NCPUS.  For GPU jobs, this should be #GPUS.
    85  * How to enforce resource share for GPUs (w/ multiple projects using GPUs)
    8663