none
Benchmark for Windows Server 2008 R2 HPC Edition (x64)

    Question

  • Здравствуйте !

    Подскажите пожалуйста, как можно протестировать кластер, собранный на Windows Server 2008 R2 HPC Edition (x64) ?

    Установил пакет Tools, там программа Lizard, запустил.  Тест проходил ОЧЕНЬ долга (реально более 3-х часов) и в результате , вместо ожидаемого 1 Тфлопса на выходе - 387 Гфлопс.

    В кластере 4 выч. ноды и 1 управляющая.  Linpack (LinX) на каждой ноде показывает около 200 Гфлопс. Почему тогда Lizard показывает такой странный результат ?

    Помогите , пожалуйста.

    Hello!

    Please tell me how to test the cluster assembled on Windows Server 2008 R2 HPC Edition (x64)?

    Installed the Tools, there is a program Lizard, launched. The test is very long (actually more than 3 hours), and as a result, instead of the expected 1 teraflops output - 387 Gflops.

    In the cluster of four Comput. nodes and one management. Linpack (LinX) on each node shows about 200 Gflops. Why then Lizard shows such a strange result?

    Help, please.

    Wednesday, December 5, 2012 6:33 AM

All replies

  • What is the theoretical throughput of you compute node?

    CPU Freq[Ghz] x cores number x number of flops op.  per core

    What result Linpack displays?

    Do you have hyperthreading on?

    What is you network bandwith beetween head node and compute node?

    Linpack may not find the best result.It is all depends on you benchmark settings mainly Ns,NBs,Ps,Qs,memory alignemnt.


    Daniel Drypczewski

    Wednesday, December 12, 2012 8:55 AM
  • What is the theoretical throughput of you compute node?

    CPU Freq[Ghz] x cores number x number of flops op.  per core

    What result Linpack displays?

    Do you have hyperthreading on?

    What is you network bandwith beetween head node and compute node?

    Linpack may not find the best result.It is all depends on you benchmark settings mainly Ns,NBs,Ps,Qs,memory alignemnt.


    Daniel Drypczewski

    Thank you very much for your reply!

    Prochessor`s- Intel E5 Xeon2660,2.2 GHz, 2 pieces percompute node, cores - 8, 8 floating point operations per clock cycle.
    The display Linpatsk(Lizard) - 500 Gflops.
    Hiperthreding turned off in the BIOS. on all the compute nodes and the control node.
    bandschith network- computer network -2Gbit/ s (test equipment cluster- 70-75MB / s

    How to make Linpak (Lizard) find the best result?
    In theory it should be about 1 teraflops, in practice - no more 600Gflops.




    Tuesday, December 18, 2012 5:49 AM
  • It looks like you got ~50% of theoretical performance.I got similiar result on my 20 node cluster with Lizard however I was able to get 65% when testing Linpack manualy (I updated hpl.dat file and adjusted a few parameters).

    Manual test on 4 nodes gave me 80% performance.Lizard best performance was 30% only.


    Daniel Drypczewski

    Thursday, December 27, 2012 9:53 AM
  • Thanks for the reply!
    But if you can detail about testing Linpack manually. Please tell me how you did it? Desirableas detailed as possible.

    Thank you in advance.

             
    Friday, December 28, 2012 5:14 AM