Thomas, On Monday 01 January 2007 08:03, Thomas Hertweck wrote:
Randall R Schulz wrote:
[...]
What sort of an application mix do you run?
We develop and run software for seismic data processing on our Linux clusters (some 64-bit, some 32-bit). In other words, it's mainly number crunching. We saw performance differences of up to 30% on the same hardware. We asked Intel about this unexpected behaviour but they've not been able to come up with an explanation yet, it's still under investigation. I agree with you that Xeon's and Core 2 Duo's characteristics might differ, but nevertheless it's an interesting observation. Unfortunately, I can't give any more details at the moment.
I take it that 30% faster is in 64-bit mode? Is that FORTRAN code? Anyway, what I need to optimize for my Core 2 Duo system is Java doing long-running symbolic (non-numeric) algorithms. I'm going to be adding concurrent processing to the system soon, now that multi-core processors are becoming common. Fortunately, many of the problems it runs are amenable to this kind of optimization on multi-core or multi-processor systems because it decomposes many problems into independently solvable sub-problems. I suppose at some point I could try benchmarking the 32/64-bit variation, but from what I've read and knowing the kind of processing my code does, I expect the advantage to remain with the 32-bit mode execution. The only thing that might be an indication for 64-bit mode is very large problems, but for this application, it's far preferable to improve the tree-search heuristics to preclude the generation of so many intermediate nodes rather than trying to accommodate them through brute force (i.e., more RAM).
Cheers, Th.
Randall Schulz -- To unsubscribe, e-mail: opensuse+unsubscribe@opensuse.org For additional commands, e-mail: opensuse+help@opensuse.org