Uploaded image for project: 'JDK'
  1. JDK
  2. JDK-4821582

3% regression in Scimark on Linux server compiler from mantis b15 to b16

    Details

    • Type: Bug
    • Status: Closed
    • Priority: P2
    • Resolution: Duplicate
    • Affects Version/s: 1.4.2
    • Fix Version/s: None
    • Component/s: hotspot
    • Subcomponent:
    • CPU:
      x86
    • OS:
      linux

      Description

      There seems to be a 3% regression in Scimark on Linux server compiler when going from 1.4.2b15 to 1.4.2b16. This has been verified on Alacrity data located at

      /java/alacrity/latest/logs/p/1.4.2+b15+l+i+s/

      and

      /java/alacrity/latest/logs/p/1.4.2+b16+l+i+s/

      You can also visualize this by going to

      http://alacrity.sfbay/query/all/nes.jsp?base=7171,7173,7174,7175,7177,7178,7179,7180,7185&build=7603,7607,7609,7549,7550,7551,7610,7611,7612&pval=0.01

      Further going into the benchmark, it looks like the subbenchmark FFT has most of the regression.

      [benkim@fastlane scimark2]$ /export/perf/b15/bin/java -server -Xmx256m -Xms256m -Xprof jnt.scimark2.commandline -large
      Composite Score: 36.58827754413559
      FFT (1048576): 14.956609377383636
      SOR (1000x1000): 85.69597360701654
      Monte Carlo : 11.869271927716074
      Sparse matmult (N=100000, nz=1000000): 27.35042835363508
      LU (1000x1000): 43.06910445492665

      java.vendor: Sun Microsystems Inc.
      java.version: 1.4.2-beta
      os.arch: i386
      os.name: Linux
      os.version: 2.4.9-31smp

      [benkim@fastlane scimark2]$ /export/perf/b16/bin/java -server -Xmx256m -Xms256m jnt.scimark2.commandline -large
      Composite Score: 35.36379994464857
      FFT (1048576): 10.254526801329945
      SOR (1000x1000): 84.7484294342001
      Monte Carlo : 11.881881015817475
      Sparse matmult (N=100000, nz=1000000): 27.444253657327685
      LU (1000x1000): 42.48990881456765

      java.vendor: Sun Microsystems Inc.
      java.version: 1.4.2-beta
      os.arch: i386
      os.name: Linux
      os.version: 2.4.9-31smp


      When running xprof on the both the versions, we see 3% more time spent in compiling FFT class.

      B15 XPROF
           Compiled + native Method
       38.3% 284 + 0 jnt.scimark2.LU.factor
       16.3% 121 + 0 jnt.scimark2.FFT.transform_internal
       13.2% 98 + 0 jnt.scimark2.MonteCarlo.integrate
       11.6% 86 + 0 jnt.scimark2.SparseCompRow.matmult
       11.5% 85 + 0 jnt.scimark2.SOR.execute
        1.9% 12 + 2 jnt.scimark2.kernel.RandomVector
        1.2% 9 + 0 jnt.scimark2.FFT.bitreverse
        1.1% 8 + 0 jnt.scimark2.kernel.RandomMatrix
        0.4% 3 + 0 jnt.scimark2.FFT.inverse
        0.3% 2 + 0 jnt.scimark2.Random.nextDouble
        0.1% 1 + 0 jnt.scimark2.kernel.CopyMatrix
       96.0% 709 + 2 Total compiled

      B16 XPROF
      Compiled + native Method
       34.1% 277 + 0 jnt.scimark2.LU.factor
       19.6% 159 + 0 jnt.scimark2.FFT.transform_internal
       15.5% 126 + 0 jnt.scimark2.MonteCarlo.integrate
       11.6% 94 + 0 jnt.scimark2.SOR.execute
       10.5% 85 + 0 jnt.scimark2.SparseCompRow.matmult
        1.7% 13 + 1 jnt.scimark2.kernel.RandomVector
        1.1% 9 + 0 jnt.scimark2.kernel.RandomMatrix
        1.1% 9 + 0 jnt.scimark2.FFT.bitreverse
        0.5% 4 + 0 jnt.scimark2.FFT.inverse
        0.1% 1 + 0 jnt.scimark2.kernel.CopyMatrix
        0.1% 1 + 0 jnt.scimark2.kernel.measureSparseMatmult
       95.8% 778 + 1 Total compiled


        Attachments

          Issue Links

            Activity

              People

              • Assignee:
                sdeversunw Steve Dever (Inactive)
                Reporter:
                duke J. Duke (Inactive)
              • Votes:
                0 Vote for this issue
                Watchers:
                0 Start watching this issue

                Dates

                • Created:
                  Updated:
                  Resolved:
                  Imported:
                  Indexed: