Uploaded image for project: 'JDK'
  1. JDK
  2. JDK-6544710

Performance regression in RangeCheckMicroBenchmark

    Details

    • Type: Bug
    • Status: Resolved
    • Priority: P4
    • Resolution: Fixed
    • Affects Version/s: 7
    • Fix Version/s: hs10
    • Component/s: hotspot
    • Labels:
      None
    • Subcomponent:
    • Introduced In Build:
      b02
    • Introduced In Version:
      7
    • Resolved In Build:
      b13
    • CPU:
      sparc
    • OS:
      solaris_9

      Backports

        Description

        Martin Buchholz wrote:

        I noticed that hotspot changes in jdk 7-b03 had dramatic
        effects on the performance of my ArrayList microbenchmark
        /net/suttles.sfbay/u/martin/ws/Iter72/test/java/util/ArrayList/RangeCheckMicroBenchmark.java


        I notice that some operations are much much faster (good!)
        but others are slower:

        mergeBench 7-b02 7-b03 jr -dsa -da -server RangeCheckMicroBenchmark.java
        ==> javac -Xlint:all RangeCheckMicroBenchmark.java
        ==> java -dsa -da -server RangeCheckMicroBenchmark
        Method Millis Ratio vs. 7-b02
        get 285 1.000 1.000
        set 185 0.651 0.995
        get/set 357 1.255 0.811
        add/remove at end 3579 12.558 1.676
        subList get 253 0.888 0.286
        subList set 187 0.656 0.177
        subList get/set 334 1.172 0.263
        subList add/remove at end 4687 16.444 1.278

        subList operations that are not structural changes are
        vastly improved, but the structural changes add/remove
        are significantly slower. (And the factor of 10 penalty
        relative to get/set seems higher than we would like)

        It would be good to know whether there is something more hotspot
        can do here. If there's interest in the server compiler
        team, I could file a bug for the add/remove performance regression.

        (I understand there are lots of performance tradeoffs here)

        Martin

          Attachments

            Issue Links

              Activity

                People

                • Assignee:
                  kvn Vladimir Kozlov
                  Reporter:
                  kvn Vladimir Kozlov
                • Votes:
                  0 Vote for this issue
                  Watchers:
                  0 Start watching this issue

                  Dates

                  • Created:
                    Updated:
                    Resolved:
                    Imported:
                    Indexed: