> performance changes in a program just based on it being run at
> different times, and getting a different set of real memory pages
> to run in.. sometimes this set of pages maps into the L2 cache
> (direct mapped) in a favorable way, the next time it maps into L2
> in a poor way.

You can get some optimisation of this btw. You can't fix interpage
clashes or crap direct map L2 caches if you have them, but Nat Friedman's
"grope" project is a tool that reorders a binary from profiling data on the
calls between functions  to put functions with a high affinity for each other
in the same page.

Reply via email to