Name: elR10090 Date: 04/15/2004


The new Hprof performance may need improvement at Windows.

I've tried it against the Context.java test (see the test's
sources in the Comments section); and the new Hprof has spent
about 20-30 minutes to write a 3-5Mb java.hprof.txt file.

Particularly, I've tried the following engineering build of
the new Hprof agent 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)

Following is a sample log I've observed for Windows 98 SE
(the machine: P-II 350MHz, RAM 384Mb, IDE 40Gb local drive):

    $ java -Xrunhprof:monitor=y Context 10 1000
    Total time (milliseconds): 19390
    Milliseconds per thread per operation: 1.939
    Dumping contended monitor usage ... done.

    real 29m 6.45s
    user 0m 0.00s
    sys 0m 0.00s

Please note, that the test itself has finished in 20 seconds;
so that almost all of these 1/2 hour were spent by the Hprof,
apparently for dumping the data into the java.hprof.txt file.

The amount of time the Hprof spends depends on a platform;
e.g., following is the log which I've seen for Solaris/SPARC
(the machine was: Ultra-10, 333MHz, RAM 256Mb, NFS drive):

    $ java -Xrunhprof:monitor=y Context 10 1000
    Total time (milliseconds): 50910
    Milliseconds per thread per operation: 5.091
    Dumping contended monitor usage ... done.

    real 1:23.2
    user 1:13.7
    sys 6.0

The difference looks dramatical -- especially taking into
account that I've tried the test using the machine's local
drive for Windows, but using an NFS drive for Solaris/SPARC.

This bug affects the following test from the NSK testbase:

    nsk/hprof/regression/Context/context007
======================================================================