Classification of ATLAS jobs by data interaction

The result should be: the average {MC/user analysis/reconstruction} job has/uses this many IOPS, input, output, storage, bandwidth.


  • look at PANDAARCH table
    • remove all the jobs where PRODUSERNAME is gangarbt (HC tests)
    • remove all the jobs where PRODSOURCELABEL is prod_test, rc_test, test, or ptest (divide on User and production jobs)
    • get number of input files, size of the input and output
  • there is no measurements of IOPS (measure it via own test jobs?)
  • Brokering jobs for FAX data access does not take into account latency (not even measured, as ping is disabled on most sites). Can it be measured?
  • Is "stagein" and "stageout" in "pilotTiming" the time it takes for the data to be transferred?

Overview: How do Cloud Providers provide Data?

  • Local Disk on VM, Block Storage ...
  • Result should be related to their pricing models.
  • Market changes quickly, results may not be valid in a couple of months


So far:

  • ran HC test jobs (Helix Nebula template) on all "Cloud" sites (specified as such in AGIS i.e. "Panda resource type": HelixNebula)
  • created HC template for Azure (Panda-Queue is available)
  • looking into FAX test jobs (by Ilija)


Goal: submit a job to a 4 core machine, that then runs simultaneously 5, 6, 7 threads to see if there is an improvement through this overprovisioning.

-- GerhardFerdinandRzehorz - 2015-05-20

Edit | Attach | Watch | Print version | History: r2 < r1 | Backlinks | Raw View | WYSIWYG | More topic actions
Topic revision: r2 - 2015-05-27 - unknown
    • Cern Search Icon Cern Search
    • TWiki Search Icon TWiki Search
    • Google Search Icon Google Search

    Sandbox All webs login

This site is powered by the TWiki collaboration platform Powered by PerlCopyright & 2008-2023 by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
or Ideas, requests, problems regarding TWiki? use Discourse or Send feedback