Messages in this thread | | | Date | Thu, 10 Jul 2008 16:58:44 -0400 | From | "H. Peter Anvin" <> | Subject | Re: [RFC 00/15] x86_64: Optimize percpu accesses |
| |
Christoph Lameter wrote: > H. Peter Anvin wrote: > >> I'm much more concerned about wasting an average of 1 MB of memory per CPU. > > Well all the memory that is now allocated via allocpercpu()s will be allocated from that 2MB segment. And cpu_alloc packs variables in a dense way. The current slab allocations try to avoid sharing cachelines which wastes lots of memory for every allocation for each and every processor. >
And how much is that, especially on *small* systems?
-hpa
| |