On 05/08/2013 09:02 AM, Mel Gorman wrote: > So preliminary testing indicates the results are mixed bag. As long as > locks are not contended, it performs fine but parallel fault testing > hits into spinlock contention on the magazine locks. A greater problem > is that because CPUs share magazines it means that the struct pages are > frequently dirtied cache lines. If CPU A frees a page to a magazine and > CPU B immediately allocates it then the cache line for the page and the > magazine bounces and this costs. It's on the TODO list to research if the > available literature has anything useful to say that does not depend on > per-cpu lists and the associated problems with them. If we don't want to bounce 'struct page' cache lines around, then we _need_ to make sure that things that don't share caches don't use the same magazine. I'm not sure there's any other way. But, that doesn't mean we have to _statically_ assign cores/thread to particular magazines. Say we had a percpu hint which points us to the last magazine we used. We always go to it first, and fall back to round-robin if our preferred one is contended. That way, if we have a mixture tasks doing heavy and light allocations, the heavy allocators will tend to "own" a magazine, and the lighter ones would gravitate to sharing one. It might be taking things too far, but we could even raise the number of magazines only when we actually *see* contention on the existing set. > 24 files changed, 571 insertions(+), 788 deletions(-) oooooooooooooooooohhhhhhhhhhhhh. The only question is how much we'll have to bloat it as we try to optimize things. :) BTW, I really like the 'magazine' name. It's not frequently used in this kind of context and it conjures up a nice mental image whether it be of stacks of periodicals or firearm ammunition clips. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>