Re: [abc] Cardinal suddenly showing strange performance results

From: Oege de Moor <Oege.de.Moor@comlab.ox.ac.uk>
Date: Sat Mar 04 2006 - 22:11:43 GMT

It seems rather unlikely that it is a problem with cardinal or the
benchmark; we have been running it for quite a while now on different
machines, with different configurations. There are quite big differences
depending on how the display is handled, but exactly as you say,
it should be faster with just vnc.

I am afraid that to include the JLO numbers in the paper at all, they
must be reproducible, so there's nothing for it but for you to
try and figure it out. Are you sure there wasn't some other major
job running on cardinal when you just tried this out?

On Sat, 4 Mar 2006, Eric Bodden wrote:

> Hello again.
>
> I just tried to track down what was causing those peeks we saw in
> yesterday's benchmark. I think I have found the reason, which was simply
> that "gc after each measurement" was _not_ enabled (I got the
> commandline arguments wrong).
>
> However, when I try to reproduce yesterday's results today, J-LO seems
> orders of magnitudes slower than yesterday! 100 iterations take multiple
> minutes already at the beginningm where yesterday they took only about
> one second! This does not depend on the fact if GC is enabled or
> disabled.
>
> And it becomes even stranger: When I run the very same code on my
> laptop, it shows results in the same order of magnitude as yesterday's
> results on cardinal - even despite the fact that on my laptop I show the
> graphics, which should impose an additional overhead. On both machines I
> am using a 1.5 JVM and 1.5 class library.
>
> I have to admit that I have no sensible explanation for this. I mean, I
> did get way faster results yesterday - you have seen the numbers. But
> there seem to be no other expensive processes running on cardinal. Also
> I have reset the vnc server. That did not help either.
>
> So I am really puzzled. I just hope that it was just me doing something
> wrong and not that it is a general problem with the benchmark. Pavel has
> helped a lot on IRC with trying to find out what could cause this but in
> the end we both went out of ideas. I really don't know what to do about
> this.
>
> Eric
>
> --
> Eric Bodden
> Sable Research Group, McGill University
> Montreal, Canada
>
>
>
Received on Sat Mar 04 22:11:44 2006

This archive was generated by hypermail 2.1.8 : Tue Mar 06 2007 - 16:13:27 GMT