質問する質問する
 

回答済みLoading XML & submitting takes lot of time

  • 2009年3月30日 7:26Lorenzo ユーザーのメダルユーザーのメダルユーザーのメダルユーザーのメダルユーザーのメダル
     
    Hello,
    We run stochastic simulations on our cluster; this kind of simulations require the submission of a lot of runs with several different parameters, passed at the command line. So, we generate and save an XML job file, load it, and submit it. With CCS 2003 loading was as fast as possible (just 1/2 seconds for a job with 30k tasks). On HPC 2008 it takes almost half an hour, on the default Job template (no additional validation required).
    Is that expected or is there something wrong with our deployment?
    Which are the exact steps going on during submission from an XML file? (does it do schema validation? other validations? Is it possible to turn them off in some config file/reg key?)
    Thanks,

    Lorenzo

回答

  • 2009年4月1日 14:55Lorenzo ユーザーのメダルユーザーのメダルユーザーのメダルユーザーのメダルユーザーのメダル
     回答済み
    They are 30000 individual tasks. We cannot use parametric tasks for this particular application (it acts on different files and parameter combinations, which can not be summarized by a single integer value).

    Lorenzo
  • 2009年5月28日 21:32Josh BarnardMSFT, 所有者ユーザーのメダルユーザーのメダルユーザーのメダルユーザーのメダルユーザーのメダル
     回答済み
    Lorenzo,
    We have some fixes around this that should be included in our first Serivce Pack; please install that SP on your client and head node when it comes out and hopefully times will be significantly reduced!

    Thanks,
    Josh
    -Josh

すべての返信

  • 2009年4月1日 13:56Steve Chilcoat - Microsoft ユーザーのメダルユーザーのメダルユーザーのメダルユーザーのメダルユーザーのメダル
     
    Lorenzo,

    Are the 30,000 tasks individual tasks or 30,000 parametric tasks?

    Thanks
  • 2009年4月1日 14:55Lorenzo ユーザーのメダルユーザーのメダルユーザーのメダルユーザーのメダルユーザーのメダル
     回答済み
    They are 30000 individual tasks. We cannot use parametric tasks for this particular application (it acts on different files and parameter combinations, which can not be summarized by a single integer value).

    Lorenzo
  • 2009年4月2日 22:03Josh BarnardMSFT, 所有者ユーザーのメダルユーザーのメダルユーザーのメダルユーザーのメダルユーザーのメダル
     

    Slow performance when adding a large number of tasks to a job is a known issue that we are looking into and hope to address soon.  In the meantime I suggest you use Parametric Sweeps as they will be much more performant to create.

    You can work around the problem you mentoined above by wrapping you task in a batch (or PSH, or ...) script that starts your application with the correct set of parameters.  You can use the * in your sweep to represent the index into the array of parameter sets, for example.

    Thanks!
    Josh


    -Josh
  • 2009年4月3日 7:15Lorenzo ユーザーのメダルユーザーのメダルユーザーのメダルユーザーのメダルユーザーのメダル
     
    Hello,

    I'd like to avoid the creation of batch; it's cumbersome and error prone.
    I suppose that we'll either write our own submission application (so that at least the user has to not wait minutes to press the "submit" button) or go back to CCS 2003 till the issue is fixed.

    Thanks!
    Lorenzo
  • 2009年4月9日 0:03Josh BarnardMSFT, 所有者ユーザーのメダルユーザーのメダルユーザーのメダルユーザーのメダルユーザーのメダル
     
    Lorenzo,
    We are working on this issue right now.  Could you please send me your XML file?   It would be helpful for us to see how it is constructed/what your job contains.

    My e-mail: JBarnard AT Microsoft DOT com

    Thanks!
    Josh
    -Josh
  • 2009年5月28日 21:32Josh BarnardMSFT, 所有者ユーザーのメダルユーザーのメダルユーザーのメダルユーザーのメダルユーザーのメダル
     回答済み
    Lorenzo,
    We have some fixes around this that should be included in our first Serivce Pack; please install that SP on your client and head node when it comes out and hopefully times will be significantly reduced!

    Thanks,
    Josh
    -Josh
  • 2009年5月29日 7:12Lorenzo ユーザーのメダルユーザーのメダルユーザーのメダルユーザーのメダルユーザーのメダル
     
    Will try it asap!
    Thank you!

    Lorenzo