Roy Lyseng wrote: > > > Victor Kirkebo wrote: >> David.Comay at Sun.COM wrote: >>>>> 1. Have you determined that -xO5 generates the best performing >>>>> code on both >>>>> SPARC and Intel. The fact that -xO5 is the highest optimization >>>>> level doesn't >>>>> always mean it generates the fastest, best performing code. >>>>> Aggressive inlining >>>>> can sometimes lead to cache thrashing, which ends up having the >>>>> exact opposite >>>>> effect from the one intended. >>>> No, I don't think we have tested this really. But I'd like to keep it >>>> for now and make a decision when we'd had time to test this. >>> In addition to the issues that Stefan brought up, higher optimization >>> levels can also lead to code generation bugs and may cause correctness >>> issues. Unless you've verified from both a correctness and performance >>> perspective that higher optimization levels actually work, you should >>> stick with the default one. >> Ok, I've removed -xO5 now. > > I cannot think that this is necessary. Surely the code will have run > through quality tests, meaning that any optimization problems should > have surfaced by now. > > The code base is also fairly small, so inlining is not likely to cause > severe cache thrashing problems. > > Obviously, compiler settings should not be changed without analyzing > performance and quality impacts, but finding the optimal settings is > probably best done at a later time. > > Roy After discussing this with the memcached team today I've put -xO5 back in. While it's true that we have not verified from a performance perspective we have done so from a correctness perspective. We have used -xO5 in both releases so far and verified correctness - we have not verified correctness without -xO5. We think it's safer to continue with -xO5 in this bug fix integration. I agree that the next full integration should verify performance as well.
Victor >> >> -Victor >>> dsc >> >> _______________________________________________ >> sfwnv-discuss mailing list >> sfwnv-discuss at opensolaris.org >> http://mail.opensolaris.org/mailman/listinfo/sfwnv-discuss -- Victor Kirkebo Database Technology Group TRO01, x43408/(+47)73842108
