General Development & Porting on Power

 View Only
  • 1.  Tuning PowerLinux for intensive scientific computing

    IBM Champion
    Posted 12 days ago
    What's the best way to learn to tune PowerLinux for intensive scientific computing ... scipy ... numpy ... etc?

    ------------------------------
    Jack Woehr
    IBM Champion 2021 -2022
    IBM Qiskit Advocate
    ------------------------------


  • 2.  RE: Tuning PowerLinux for intensive scientific computing

    Posted 12 days ago

    Jack,

    Let me give you 2 simple tricks and include a link for more detailed information.

    • First trick:  scipy and numpy code often ran single threaded, thus they can be very sensitive to the SMT setting on POWER.   By default the KVM capable AC922 & IC922 systems run in SMT4 mode, while PowerVM capable systems like S922s, E950, E980 and S1022, E1050 and E1080s run in SMT8 mode.  To change the SMT level in PLinux, run the following command as root: 
      • To find out what SMT level you are using right now:
        ppc64_cpu --SMT
      • To set the SMT level to 4 or 2
        ppc64_cpu --SMT=4
        ppc64_cpu --SMT=2
    • Second trick: use python libraries that are optimized for the POWER platform.   For example if you are using conda you can use the rocketce channel
      conda config -add channels rocketce​​

    Here is a paper which provides more performance information and tips: https://developer.ibm.com/tutorials/migrate-app-on-lop/



    ------------------------------
    Theresa Xu
    ------------------------------



  • 3.  RE: Tuning PowerLinux for intensive scientific computing

    Posted 11 days ago
    I'll sneak in here with a minor correction:  the "--SMT" parameter is actually lower case "--smt".

    ------------------------------
    Paul Clarke
    ------------------------------



  • 4.  RE: Tuning PowerLinux for intensive scientific computing

    IBM Champion
    Posted 11 days ago
    Thanks Paul!

    ------------------------------
    Jack Woehr
    IBM Champion 2021 -2022
    IBM Qiskit Advocate
    ------------------------------



  • 5.  RE: Tuning PowerLinux for intensive scientific computing

    IBM Champion
    Posted 11 days ago
    Thanks Ms Xu!

    ------------------------------
    Jack Woehr
    IBM Champion 2021 -2022
    IBM Qiskit Advocate
    ------------------------------



  • 6.  RE: Tuning PowerLinux for intensive scientific computing

    Posted 11 days ago
    re: rocketce -- Here is the link to information about their latest releases of Power-optimized builds:

    https://community.rocketsoftware.com/discussion/rocketce-releases

    Definitely read through them -- some of the packages require specific versions on the install (ie, pytorch-cpu=1.10.2) in order to get the 'best on power' version and not have it upgraded to a newer (but slower) version from somewhere else.

    .bri.

    ------------------------------
    Brian Horton
    ------------------------------



  • 7.  RE: Tuning PowerLinux for intensive scientific computing

    Posted 11 days ago
    Edited by Brian Horton 11 days ago