3 Replies Latest reply on Mar 26, 2014 7:26 AM by kevin_intel

    performance variation on Sandybridge (Xeon E5-2680), why?

    merryhy

      Dear all,

       

      We get a issue from our customer, one of their test program’s performance varied from run to run on one Sandybridge 2-way server. For example, one test case runs sometimes 12.35s, sometimes 9.32s. 

      The test exe is very simple, just do matrix multiply.  I upload the exe and cpp file at http://software.intel.com/file/44733  and  http://software.intel.com/file/44734

       

      What we have known,

      the user's system: please see the cpuz result as cpuz.png,  Intel Xeon E5 2680,  2 processor, each processor 8 cores.

        Turbo Boost is Disabled,

              C state is disabled.

              Hyper-threading is disabled.

              Microsoft Windows Server 2008 R2 Standard installed , Memory 32728MB (32GB)  , c: 690GB.

               The test machine almost don't run any other process except OS and anti-virus process.

       

              We asked the user to try

              disable all of other OS processs and ran the TestBLAS.exe with anti-virus active/inactive, same result.

           CPU usage: At no time during the TestBLAS runs did the CPU % get above 6% while TestBLAS was running that is according to Task Manager.

           We also  ask the user to try the affinity setting ( START  /B /AFFINITY 0x000000FF testblas), same result.cpuz.png

      I have almost same machine, but i tried the TestBlas.exe many times and always see the same timing.  This seems to be an issue related to the user’s system. But no clue,  is it could be known issue of sandybridge or cpu tempeture issue?

       

      a    Any idea is appreciated!

       

      Th     Thanks

              Ying

       

        • 1. Re: performance variation on Sandybridge (Xeon E5-2680), why?
          Adolfo_Intel

          Since you are unable to replicate the issue on your own system, it seems that it is a behavior happening only on your customer system.

          So I would suggest performing more tested on your customer system, like running with less memory RAM, or making sure no RAID configurations are available.

          Check if the processor on your system is the same model of the one on your customer system, since different processor might have different behaviors or results.

          • 2. Re: performance variation on Sandybridge (Xeon E5-2680), why?
            philipw

            I am having the same issue on 3 different servers where we're looping 20,40,60 times on a compression of an image file.  all three servers have sandy bridge xeon processors, same generation as OP.  We get inconsistent readings, going from 40ms (optimal) to 80ms spikes throughout the test.  2 of these are running on physical hardware and one is on virtual machine.

             

            These same test ran on older generation Xeons and i7 processors doesn't run into this issue, consistently getting the fastest time through the whole test.  One thing I did notice is the Sandy Bridge processors are stuck at 1100mhz, and rarely shoot up to it's max frequency.  I have tried booting with turbo boost, c1e and c-state off, and it does not force the CPU to run at max frequency.  There is a serious performance hit with these processors and i can't figure out why.  On an older server with a xeon x5000 series with c1e, turbo boost and c-state on, it stays running at 3.5ghz and ran the tests flawlessly. 

            • 3. Re: performance variation on Sandybridge (Xeon E5-2680), why?
              kevin_intel

              Philipw, I would recommend you posting your query on the Server Community for them to provide assistance on your systems.

               

              Here is the community:

              https://communities.intel.com/community/tech/servers