Name: elR10090 Date: 04/15/2004


The new Hprof performance may need improvement.

I've tried the HeapThrasher.java test (see the test's sources
in the Comments section); and test worked about 10 times slower
with -Xrunhprof option rather than without the new Hprof agent.

Particularly, I've tried the following engineering build equipped
with the libhprof.so (hprof.dll) libraries provided by Kelly:

    java version "1.5.0-internal"
    Java(TM) 2 Runtime Environment, Standard Edition (build
1.5.0-internal-ohair_06_apr_2004_17_28)
    Java HotSpot(TM) Server VM (build 1.5-internal, mixed mode)

The performance gap is observed for all platforms I've tried:
- Solaris 9: 2xCPU SPARC 400MHz, RAM 1024Mb, HDD SCSI 36Gb
- Solaris 8: 1XCPU P-III 866MHz, RAM 512Mb, HDD IDE 80Gb
- Linux RH9: 2XCPU P-III 600MHz, RAM 512Mb, HDD IDE 40Gb
- Win2K : 2xCPU P-III 600MHz, RAM 384Mb, HDD SCSI 10Gb
- Win98 SE : 1xCPU P-II 350MHz, RAM 384Mb, HDD IDE 10Gb
(all runs were made using the machine's local HDD).

Following is the sample log for the worst case I've observed.
That was about 11 times performance gap for the Win2K platform:

    $ java HeapThrasher 2 2
    Thu Apr 15 18:59:57 NOVST 2004
    Allocated: 1250000
    Freed: 0
    Total Memory: 66650112
    Free Memory: 8514784
    Completed iteration 0
    Thu Apr 15 19:00:30 NOVST 2004
    Allocated: 6499000
    Freed: 5862000
    Total Memory: 66650112
    Free Memory: 21762232
    Completed iteration 1

    real 0m58.85s
    user 1m12.03s
    sys 0m 3.94s


    $ java -Xrunhprof HeapThrasher 2 2
    Thu Apr 15 19:02:57 NOVST 2004
    Allocated: 1250000
    Freed: 250000
    Total Memory: 66650112
    Free Memory: 10188008
    Completed iteration 0
    Thu Apr 15 19:03:39 NOVST 2004
    Allocated: 1500000
    Freed: 1250000
    Total Memory: 66650112
    Free Memory: 2614456
    Completed iteration 1
    Dumping Java heap ... allocation sites ... done.

    real 11m 8.17s
    user 6m55.28s
    sys 1m12.17s

This issue affects the following test from the NSK testbase:

    nsk/hprof/regression/HeapThrasher

The test will fail if 10-times or worse performance gap is
observed.
======================================================================

Name: dd4877 Date: 04/23/2004


daniel.daugherty@Sun 2004-04-23

This bug also affects the following test:

    nsk/hprof/regression/b4478223



======================================================================