From mboxrd@z Thu Jan 1 00:00:00 1970 Date: Tue, 7 Nov 2006 16:30:28 +0000 (GMT) From: Mel Gorman Subject: Re: Page allocator: Single Zone optimizations In-Reply-To: Message-ID: References: <4544914F.3000502@yahoo.com.au> <20061101182605.GC27386@skynet.ie> <20061101123451.3fd6cfa4.akpm@osdl.org> <454A2CE5.6080003@shadowen.org> MIME-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII; format=flowed Sender: owner-linux-mm@kvack.org Return-Path: To: Christoph Lameter Cc: Andy Whitcroft , Andrew Morton , Nick Piggin , KAMEZAWA Hiroyuki , Linux Memory Management List , Peter Zijlstra List-ID: On Fri, 3 Nov 2006, Christoph Lameter wrote: > On Fri, 3 Nov 2006, Mel Gorman wrote: > > > > I think we need something like what is done here via anti-frag but I > > > wish > > > it would be more generic and not solely rely on reclaim to get pages > > > freed > > > up. > > > > > > > How could it have been made more generic? Fundamentally, all we are > > doing at > > the moment is using the freelists to cluster types of pages together. We > > only > > depend on reclaim now. If we get the clustering part done, I can start > > working > > on the page migration part. > > Right lets have a special freelist for unreclaim/unmovable pages. I think > that we agree on m that. Somehow we need to be able to insure that > unXXXable pages do not end up in sections of the zone where we allow > memory hotplug. > Ok, I redid the patches (calling it Page Clustering to distinguish between anti-frag) to split between two types of pages o Unmovable and unreclaimable pages o Movable pages and those slab pages that are reclaimable The per-cpu allocator could only be used for the second type of pages (i.e. Movable/Reclaimable pages) to avoid leaking or expanding the per-cpu structures. I thought it would work ok but have a tendancy to reap caches aggressively to satisfy hugepage allocations. I was wrong, it's success rates for hugepage allocations was even lower than the normal allocator on ppc64. On x86_64, it was a bit better than the standard allocator, probably because of the smaller hugepage size, but nowhere near as successful as anti-fragmentation. To guess what went wrong, I examined the slabinfo on ppc64 after the tests, I found that even though slab was reaped aggressively, 105 pages were still in use by the ext3 inode cache. As there are only 62 hugepages on the system, it's easy to see why they would end up in awkward locations. Here are some results at the end of tests when the system is back at rest. They are the percentage of memory that can be allocated as hugepages after a series of tests that cause fragmentation. The system is a ppc64 with 1GiB of RAM Standard allocator: 9% Page Clustering: 3% Anti-fragmentation: 56% Here are the results on x86_64 which has a much smaller hugepage size Standard allocator: 21% Page Clustering: 33% Anti-fragmentation: 56% I know from previous experience that keeping slab pages with unmovable pages does not work very well because updatedb or memory pressure tends to cause problems. Hence, I'm still convinced that slab pages for caches like inode and short-lived allocations need to be clustered separetly. > At some later point we would like to have the ability to redirect > unXXXable allocations to another node if the node is hot pluggable. > That can be done separately. > > > Also the duplication of the page struct caches worries me because it > > > reduces the hit rate. > > > > do you mean the per-cpu caches? If so, without clustering in the per-cpu > > caches, unmovable allocations would "leak" into blocks used for movable > > allocations. > > I mean the per cpu caches and I think you could just bypass the per cpu > caches for unXXXable pages. Kernel pages are buffered already in the slab > allocator and other kernel allocations are probably rare enough. > I'll revisit the patches again, cluster slab caches separately but only allow movable allocations to go through per-cpu and see what happens. > > > Removing the intermediate type would reduce the page > > > caches to 2. > > > > And significantly reduce the effectiveness of the clustering in the > > process. > > Are you sure about this? I think the figures support me on this. > It seems that the intermediate type is > reclaimable and you already allow "reclaimable" pages to be not > reclaimable > (mlock'ed pages). I know, but not because I'm happy about it. As mlock() can be called after a fault, knowledge of the future is required to keep all mlocked pages out of "reclaimable" areas. It's similar for pages pinned by RDMA. To address mlock() pages properly, page migration needs to be used. I can create a patch later to handle the case where a page is being faulted in due to mlock() and not to mark it "reclaimable" if it's known that they are the majority of mlocked pages. > If you run into trouble with the reclaimable slab pages > in the reclaimable zone then you could do agressive slab reclaim to remedy > the situation. > That made intuitive sense so I checked it out but doesn't work in practice even when slabs are aggressively reaped. > > > And maybe we do not need caches for unreclaimable/unmovable > > > pages? slab already does its own buffering there. > > That is true. If it is a problem, what could be done is have a per-cpu > > cache > > for movable and unmovable allocations. Then have the __GFP_KERNRCLM > > allocations bypass the per-cpu allocator altogether and go straight to > > the > > buddy allocator. > > Right. Maybe we can get away with leaving the pageset cpu caches > untouched? On our largest systems with 1k nodes 4k cpus we currently have > 4 zones * 4096 cpus * 1024 nodes = 16 million pagesets. Each of those has > hot and cold yielding 32 million lists. Now we going triplicate that to > 192 mio lists and we also increase the size of the structure. > I can see the problem with expanding the per-cpu structures. I'll check out what happens when per-cpu caches are only used for movable allocations. This is the way things were in an earlier version of anti-fragmentation but I do not have figures any more. > With the code currently in 2.6.19 we go from 4 to 2 zones. So we have only > 16 million pagesets. With the optional DMA in mm we got from 16 to 8 > million pagesets. This effectively undoes the optimizations done in .19 > .20. > -- Mel Gorman Part-time Phd Student Linux Technology Center University of Limerick IBM Dublin Software Lab -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@kvack.org. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: email@kvack.org