Messages in this thread | | | Date | Wed, 30 May 2012 05:41:16 -0400 | From | KOSAKI Motohiro <> | Subject | Re: [PATCH 13/35] autonuma: add page structure fields |
| |
(5/30/12 5:06 AM), Peter Zijlstra wrote: > On Wed, 2012-05-30 at 04:25 -0400, KOSAKI Motohiro wrote: >> (5/29/12 10:54 AM), Peter Zijlstra wrote: >>> On Tue, 2012-05-29 at 09:56 -0400, Rik van Riel wrote: >>>> On 05/29/2012 09:16 AM, Peter Zijlstra wrote: >>>>> On Fri, 2012-05-25 at 19:02 +0200, Andrea Arcangeli wrote: >>>> >>>>> 24 bytes per page.. or ~0.6% of memory gone. This is far too great a >>>>> price to pay. >>>>> >>>>> At LSF/MM Rik already suggested you limit the number of pages that can >>>>> be migrated concurrently and use this to move the extra list_head out of >>>>> struct page and into a smaller amount of extra structures, reducing the >>>>> total overhead. >>>> >>>> For THP, we should be able to track this NUMA info on a >>>> 2MB page granularity. >>> >>> Yeah, but that's another x86-only feature, _IF_ we're going to do this >>> it must be done for all archs that have CONFIG_NUMA, thus we're stuck >>> with 4k (or other base page size). >> >> Even if THP=n, we don't need 4k granularity. All modern malloc implementation have >> per-thread heap (e.g. glibc call it as arena) and it is usually 1-8MB size. So, if >> it is larger than 2MB, we can always use per-pmd tracking. iow, memory consumption >> reduce to 1/512. > > Yes, and we all know objects allocated in one thread are never shared > with other threads.. the producer-consumer pattern seems fairly popular > and will destroy your argument.
THP also strike producer-consumer pattern. But, as far as I know, people haven't observed significant performance degression. thus I _guessed_ performance critical producer-consumer pattern is rare. Just guess.
>> My suggestion is, track per-pmd (i.e. 2M size) granularity and fix glibc too (current >> glibc malloc has dynamically arena size adjusting feature and then it often become >> less than 2M). > > The trouble with making this per pmd is that you then get the false > sharing per pmd, so if there's shared data on the 2m page you'll not > know where to put it. > > I also know of some folks who did a strict per-cpu allocator based on > some kernel patches I hope to see posted sometime soon. This because if > you have many more threads than cpus the wasted space in your areas is > tremendous.
| |