locked
Incorrect number of cores identified with Windows Server HPC/Xeon 5550 RRS feed

  • Question

  • I have two IBM x3650 M2 servers, each with 2 Xeon 5550 quad core cpus.

    I have installed the HPC version of Windows Server 2008 and setup the HPC cluster.

    On each server the number of cores identified is 16, double what is actually there.

    I have installed all the latest updates and used the latest IBM server setup utility and still have the problem.

    I have tried disabled hyper-threading, but the HPC manager is still identifying 16 cores per machine. (this also slows down the simulations dramatically).

    Wednesday, June 30, 2010 11:05 AM

Answers

  • Hi there

    Now that you've disabled hyper-threading, is the OS reporting the correct number of cores? If so, please use HPC manager to take each node offline, then back online. This process should ensure that the cluster has correct configuration knowledge of the nodes in question.

    Cheers

    Dan

    Wednesday, June 30, 2010 12:24 PM

All replies

  • Hi there

    Now that you've disabled hyper-threading, is the OS reporting the correct number of cores? If so, please use HPC manager to take each node offline, then back online. This process should ensure that the cluster has correct configuration knowledge of the nodes in question.

    Cheers

    Dan

    Wednesday, June 30, 2010 12:24 PM
  • Hi,

    Disabling HT and then taking the nodes off-line and then back on does result in the HPC manager recognising 8 cores per server.

    However comparing a few benchmarks shows that disabling HT slows down the simulations (ANSYS CFD) by 20 to 40%.

    The main problem here is that inorder to make full use of the computing capacity I have to start jobs over "32 cores" on the two servers which takes up twice the ANSYS licenses (which are issued on a per core basis).

    Sunday, July 4, 2010 10:22 PM
  • So, let me see if I understand. You have 2 nodes, each with 2x4 core processors, so 16 physical cores in total. Your jobs are Ansys simulations, which you naturally want to run as efficiently as possible. Your benchmarks show a 20-40% increase in performance when run 32 way over hyperthreaded processors.

    I don't think this is a problem as such, but is more about deciding on how to use your cluster in the most effective way for your simulations and environment. If, as you say, you enable hyper-threading, and run your simulations 32 way, you achieve the best performance, the compromise is Ansys license cost. If you turn off hyper-threading, you lose 20-40% performance, but do not take the license cost hit.

    If you run a 32 way job over 16 cores without HT turned on you will see horrible performance!

    What performance do you see comparing a 16 way job over HT processors Vs a 16 way job over non HT processors? Could you make use of hyper-threading, but run 16 (or 24) way and use the rest of the cores for other operations (e.g. pre / post processing)?

    Regards

    Dan

     

    Tuesday, July 6, 2010 8:04 AM