Why is SciMark benchmark test so low on my machine ?

[Pascal]

Hi all,

I have to start doing some numerical computations in Java. To check
Java's performance on my computer, I have run the Scimark benchmark
(http://math.nist.gov/scimark2/) test. The results put me somewhere at
the performance of the Sun 1.1.8 JVM on 1.5-2.0 Ghz boxes. I used Sun
1.5.0 beta2 and Sun 1.4.2 03 JVMs on a P4 2.8 GHz machine and a P
Xenon 2.4 GHz Dual Processor machine. Comparable configurations were
three times as fast (as indicated by the results page) as mine.

Is there any way to tune the JVM ? I 've played around with several
switches, but the results are always poor. I have also shut down
several background process which could slow down JVM. No performance
increase.

Any help would be highly appreciated.

Pascal

 

[Paul Lutus]

Hi all,

I have to start doing some numerical computations in Java. To check
Java's performance on my computer, I have run the Scimark benchmark
(http://math.nist.gov/scimark2/) test. The results put me somewhere at
the performance of the Sun 1.1.8 JVM on 1.5-2.0 Ghz boxes. I used Sun
1.5.0 beta2 and Sun 1.4.2 03 JVMs on a P4 2.8 GHz machine and a P
Xenon 2.4 GHz Dual Processor machine. Comparable configurations were
three times as fast (as indicated by the results page) as mine.

How much memory do you have? You left this out. How fast is your hard drive
(if a lot of data are being handled)? What operating system? You left this
out.

Is there any way to tune the JVM ?

If you haven't already, you can always try:

java -server (classname)

I 've played around with several
switches,

Which ones?

 

[Pascal]

Thanks Paul, "-server" tripled performance. I was not aware of
different JVMs. Well, just moving from Fortran to Java ...

Pascal

 

[Sudsy]

Thanks Paul, "-server" tripled performance. I was not aware of
different JVMs. Well, just moving from Fortran to Java ...

Pascal

Paul deserves kudos for identifying the problem and suggesting a
practical solution! We're all in this to learn and share experiences
and I frequently read innovative and new (to me) approaches.
I just get bummed-out sometimes when people complain that Java
doesn't do it like "X" (insert pet language here).
I view the Java platform as the ultimate LEGO(tm) block set. You
can find almost anything you need to build upon. It's not perfect,
but you can implement interfaces or extend classes to your heart's
content and construct industrial-strength solutions.
YMMV

 

[Chris Uppal]

Thanks Paul, "-server" tripled performance. I was not aware of
different JVMs. Well, just moving from Fortran to Java ...

I would expect that -server would increase the speed of heavy arithmetic code
since the server JVM puts more emphasis on "advanced" optimisations in the
JITer. But before reading too much into the result you should check that the
optimiser hasn't just optimised some of the test code away entirely. It is
quite capable of recognising that some loop has no side-effects and removing it
(I had to re-write one micro-benchmark of my own for just this reason).

I haven't looked at the SciMark benchmarks (yet), and it may be that this is
not an issue for them, but I think it's worth checking.

BTW, I get strange results on the SciMark benchmark too (on two machines, Win2K
and WinXP, running 1.5 beta1 and 1.4.1 respectively). Without investigating
further, I would not be inclined to put much credence in SciMark. I wouldn't
/expect/ NIST to ****-up such a thing, so my first guess is that something may
be screwing up the reported results.

-- chris