Ruud Schramp <spam2@xxxxxxxxxxx> wrote: > > you refer to schedulers/os/compilers mostly closed source and make > some general remarks about schedulers. No, the remarks I made were independent of whether the code is open or closed source. > If applies do Gcc and linux 2.6 how do they turn out. I haven't, I am afraid, but I have reasons to believe that my post applies to them as well. A couple of decades ago, I thought of the problem "How would I design a language like OpenMP to enable mere mortals to write efficient code?" I am still thinking :-) My current conclusion is that it can't be done without a major rethink of the principles of hardware and operating systems. The problems are deeper than the language. > As you can see in the original email, the inner loop hardly uses any > data from the global data. I would expect no significant overhead by > cache or memory by that argument. Don't bet on it! You need to go a LOT deeper into the memory subsystem than THAT. In particular, you need to study cache, memory and often TLB associativity. Look up 'false sharing' on the net. > Hmmm, a carreer in low level stuff. I love it ;) The best of luck to you - I am approaching retirement, and I can tell you that there are virtually no people in the UK with serious OpenMP tuning skills. A very few companies (like NAG) have a couple, and the occasional university has the odd one (usually a VERY odd one!), but that is all. Regards, Nick Maclaren, University of Cambridge Computing Service, New Museums Site, Pembroke Street, Cambridge CB2 3QH, England. Email: nmm1@xxxxxxxxx Tel.: +44 1223 334761 Fax: +44 1223 334679
