wiki:Discussion-08.06.2008
Last modified 9 years ago Last modified on 08/07/08 09:49:02

This page records discussion we had on leveraging ASA resources more efficiently

  • We started out with, if the other 2 ASA resources Cray XD1 and DMC, in addition to Altix can be included individually in the IM-MAD of Gridway, so that we can target our jobs directly to these resources from Gridway.
  • Not possible since, Altix is the only ASA resource which has Globus installed. Cray XD1 and DMc do not have Globus installed.
  • The reason we were pondering on the above inclusion, was that DMC offers so much faster execution than either Cheaha/Olympus?.
  • For a 100 iteration job, where Cheaha/Olympus? take > 1 hour to finish, it takes only 25 minutes for the same 100 iteration to complete on DMC.
  • DMC runs on 3 GHz clock. Olympus though runs on a 3 GHz clock too, it does not offer the speed as DMC. Three factors were assumed to contribute to the faster execution on DMC
    1. The memory architecture/configuration on DMC which makes it execute faster
    2. The compiler on DMC might also make a difference
    3. The installed software version of R-statistical language on DMC.
      ASA recently installed R-2.7.1 version on its resources. The highest version of R available on Cheaha and Olympus is 2.6.2.
      Tickets: #24 Install R-2.7.1 in $HOME dir on Cheaha and Olympus, execute jobs and see if there's any improvement in the performance.
  • Next, we looked at how we can harness DMC's computing power from Gridway.
  • John-Paul said that if Gridway has a substitution variable like ${GW_HOSTNAME}, we can send a separate executable for each host
    Tickets: #25 I think, though I didn't try this, can be done, because the variable ${GW_HOSTNAME} does appear on the target resource, when printing out the env variables by the executable
    Tickets: #26 The above variable, assuming to not exist, struck upon the idea that, inside our executable which will be submitted from Gridway, if we mention a PBS directive to target the job to execute on DMC, then we might speed up our performance.
    Tickets: #27 Additionally, verify how long would a single chunk of 1000 iteration job takes to complete on DMC