Re: CPU affinity & IPI latency

[Mark Hahn]

> If the amount of data/instructions needed by all 5 processes fits in the 
> cache and you pin all the processes to the same CPU you'll get much 
> better performance than simply letting them float.

interesting.  I remember that around 2.3.40, the scheduler handled
"frequent-schedulers" (like lat_ctx) differently, based on how long 
a timeslice the proc used, relative to the estimated time to flush cache.

as I recall, it let them stay on their current CPU.  like letting 
someone with 1 item go ahead of you in a grocery store checkout ;)

that code is mostly gone (only cacheflush_time remains);  I think it
morphed into the current migrate-to-longest-idle heuristic.

does anyone remember why the frequent-schedulers code was killed?
just because it conflated cache-affinity with timeslice?

regards, mark hahn.