Hi.
> > Let me elaborate:
> > At some point of the simulation, the same quantity Q can be computed in 2
> > different ways, say "QH" and "QL". When using "QL", the program is about
> > twice as fast as when using "QL". This is confirmed when using "-Xint".
> > Leaving out "-Xint" (for maximum performance), the result is, quite
> > surprisingly, the other way around: Using "QL" is about 50% slower than
> > with "QH"!
> >
> > Doesn't this indicate some subtle bug in the JIT compiler?
>
> What about the absolute performance numbers?
I'm not sure I understand what numbers you are interested in.
Here are some (approximate) numbers (time in minutes) to give an idea
of the problem:
QL | QH
--------+-------
1.6 -Xint 60 | 120
1.6 JIT 15 | 10 <---- Something wrong here, I think.
1.5 JIT 6 | 12
> As long as long-running
> code doesn't run slower without -Xint, I don't think this is a JIT
> bug.
Well, the fully interpreted program is intolerably slow of course.
But does that mean that one should be happy with whatever speedup is
achieved, even if the objective relative performance of algorithms is
reversed?
> It's somewhat expected that relative performance changes.
How could that be expected???
As I indicated in my report (see also the last line in the above table), the
algorithms ("QL" vs "QH") relative performance is consistent when using java
1.5; it is not with java 1.6 ...
Going from 1.5 to 1.6, I got about 20% performance increase when the code
uses "QH". So, I would expect some gain (however small) for "QL" too. What
I would certainly not expect is a 250% loss!
I hope that this was clearer than my first explanation.
Best regards,
Gilles
--
To UNSUBSCRIBE, email to [EMAIL PROTECTED]
with a subject of "unsubscribe". Trouble? Contact [EMAIL PROTECTED]
