none
Lizard benchmark tool, runtime? RRS feed

  • Question

  • Hi.

    I was woundering about this benchmarking tool Lizard that use Linpack. How long time takes it to run Lizard? It have been running for a day now and still running. I choose Advanced Tuning Mode because in Express Tuning Mode it coulnt find any nodes. I have not made any changes to the parameters. In node management the head nodes CPU is 100% is that normal.

    Any one out there that have experience with Lizard? or should i go for Linpack only?

    My setup is 2 workstation nodes and 1 headnode.

    Best regards /Marcus 


    Zorky HPC

    Friday, March 16, 2012 10:05 AM

All replies

  • Hello.

    So sad, but i had similar experience using lizard. I know that it trying to tweak its algorithm for achieve best result on the architecture on which it is running, but i waited for three days and it continued to tweak. I suppose, that you can stop calculations and study intermediate results.

    But it's not normal that lizard can't find any nodes in Express Tuning Mode. Please share your configuration with us and how you installed your cluster - maybe, somewhere error is lurking.

    And yes, try to use just linpack.

    Friday, March 16, 2012 12:40 PM
  • Hi Marcus, I'm pretty sure Lizard only works on nodes in the 'compute nodes' group and not in the 'workstation nodes' group. Tuning workstation nodes wouldn't be a core scenario, since that implies you are running on user's desktops and across your corporate intranet, which would add two large variables to your runs.
    Sunday, March 25, 2012 5:58 AM
    Moderator
  • Hi Don.

    That i want to do is to measure the hole cluster capacity. This is just a test scenario, no users are working on the worksations. And my setup is just a headnode whit 2 workstation (for now) isolated from my corporate network.

    Best regards Marcus


    Zorky HPC

    Monday, March 26, 2012 6:40 AM
  • Hi,

    in that case I think you'll be better off deploying these workstations as compute nodes.

    Michael

    Tuesday, March 27, 2012 11:52 PM
  • Hi Michael.

    Yes, it would be better with compute nodes. But im trying to compare them, compute node vs workstation.

    Best regards Marcus


    Zorky HPC

    Wednesday, March 28, 2012 5:02 AM
  • Hi Zorky,

    Workstation node has the same binary as compute nodes, with a dedicated service watching for user interaction, so that if a user wants to use workstation node, the HPC job will not be scheduled to that node. The purpose of creating workstation node is to let the cluster user utilize cpu cycles when their office machines are at idle time (say, at night), there's really no loss/gain when you compare compute node and workstation node.

    Michael

    Wednesday, March 28, 2012 7:23 AM