Performance comparision of a XP1000 6/500 and GS140 6/525

From: Uwe Lienig (uwe.lienig_at_fif.mw.htw-dresden.de)
Date: 12/23/04

  • Next message: Marquess, Dustin: "SUMMAY: Stable version of GNU screen"
    Date: Thu, 23 Dec 2004 13:24:14 +0100
    To: Tru64 UNIX <tru64-unix-managers@ornl.gov>
    
    

    Hi managers,

    merry christmas to all, whoever will celebrate it.

    I've come across the following performance question:

    I have a XP1000 6/500, 2GB mem, OS v5.1 and a GS140 6/525, 8 CPU, 8GB mem
    (1x4GB, 2x2GB), OS v5.1b.

    The GS140 is an upgraded 8400 that now calls itself a GS140 (I did the upgrade
    from a AS8400 5/300 to a AS8400 6/525 by replacing the required CPU modules
    E2054 -> E2063 and other, now the console says GS140).

    The GS140 is dedicated to number crunching of medium sized CFD modells (up to
    256MB data). The CFD code can not be run multithreaded or in a MPI
    environment. Since the number of simulations to be carried out exceeds 20 I
    decided to run 8 simulation simultaneously on the GS140, so that the total
    amount of time that is required is acceptable. Instead of having solved one
    simulation in a short time with 2 or 4 CPU's I'll get 8 simulations
    calculated at once.

    Since the GS140 has plenty of mem (for 8 simulations I'll need something
    around of 4 GB) and I will make sure that there will be running only one
    simulation per CPU I expected no performance degradation.

    Since the XP1000 is equipped with nearly the same CPU as the GS140 I checked
    the run time of the same modell on the XP1000 while the GS140 was loaded with
    7 other simulations (a total of 8 on the GS140). I expected that the run time
    of the simulation on the GS140 may be a little bit higher when compared to
    the XP1000 (mem access of one CPU on the GS140 may be delayed due to
    concurrent access of another CPU). But the delay shouldn't increase the
    processing time considerably.

    But to my surprise the GS140 needed 4 times the amount of processing time
    compared to the XP1000. I was astonished that the processing time increased
    that much. Right after having the GS140 up and running I loaded only one
    simulation on the GS140 and I seem to remember that the processing time was
    comparable to the XP1000 then. Right now there are 8 simulations running and
    I'm reluctant to kill them, even though they aren't performing optimal,
    because the simulations will end by the weekend. Furthermore I'm not certain
    what measures to use to increase the performance of the GS140,

    I'm not new to Tru64 but fairly new to the midrange AlphaServers (8400/GS) and
    have no experience what may be expected in the situation I'm facing here.
    My questions are related to the performance:

    Do I have to live with this performace decay, when the GS is fully loaded with
    these CPU and mem demanding processes?

    Will simultaneous access of the CPU's to the mem decrease the performance that
    much?

    If I use collect or the other tools may the performance be increased
    considerably?

    TIA

    -- 
    Uwe Lienig
    ----------
    fon: (+49 351) 462 2780
    fax: (+49 351) 462 3476
    mailto:uwe.lienig@fif.mw.htw-dresden.de
    Forschungsinstitut Fahrzeugtechnik
    <http://www.fif.mw.htw-dresden.de>
    parcels: Gutzkowstr. 22, 01069 Dresden 
    letters: PF 12 07 01,    01008 Dresden
    Hochschule für Technik und Wirtschaft Dresden (FH)
    Friedrich-List-Platz 1, 01069 Dresden
    

  • Next message: Marquess, Dustin: "SUMMAY: Stable version of GNU screen"

    Relevant Pages

    • Re: Slower performance in ActiveXscript task
      ... the processing time increased to 2 hour. ... grateful if anyone can help for such strange case in DTS task. ... using CPU - ActiveX can use a lot of cpu. ... Or, did you switch from ...
      (microsoft.public.sqlserver.dts)
    • Re: FPGA tool benchmarks on Linux systems
      ... because the other application will run from the other CPU. ... > If I interpret the data correctly, two CPU result in the same simulation ... >> I had only one not so state of the art Opteron to measure. ... >> RTL Both CPUs active ...
      (comp.os.linux.hardware)
    • Re: Java Prozesse zu langsam, Leerlaufprozess > 90 %
      ... Ich habe eine Simulation unter Java runtime environment laufen. ... Konsolen App, sonst mit SendSignal.exe) und schauen, was Deine App intern ... dass der Prozess keine CPU anfordert. ...
      (de.comp.lang.java)
    • Re: FPGA tool benchmarks on Linux systems
      ... If I interpret the data correctly, two CPU result in the same simulation ... > I have some measurements with more current Xeon processors. ... > RTL Both CPUs active ...
      (comp.os.linux.hardware)
    • Re: Java Prozesse zu langsam, Leerlaufprozess > 90 %
      ... Ich habe eine Simulation unter Java runtime environment laufen. ... Konsolen App, sonst mit SendSignal.exe) und schauen, was Deine App intern ... dass der Prozess keine CPU anfordert. ...
      (de.comp.lang.java)