On Fri, 12 Sep 2008 19:06:49 +0530 "amey mahamuni" <ameymahamuni@xxxxxxxxx> wrote: > I am ameya and I want to do a 3-4 month project in linux memory > management > I am interested in page replacement policy.as I am new to linux kernel I not > able to select particular page replacement algorithm for study. > so, I would like to know any enhancement or any improvement in present > algorithm . what page replacement algorithm is used by latest kernel version The latest kernel from Linus uses a complex combination of replacement algorithms. There are two pageout lists (active and inactive), with different policies for pages, depending on whether they are swap backed, anonymous, filesystem backed and mapped or unmapped. The -mm tree has the split LRU VM, which has more pageout lists: active anon, inactive anon, active file, inactive file and unevictable. Anonymous (and other swap backed) pages use a simple two-handed clock replacement algorithm, while file backed pages use a use-once algorithm that somewhat resembles 2Q. Pages that cannot be evicted are placed on the unevictable list, to keep them out of the way of the VM. The split LRU VM has a fairly simple balancing scheme to put different pageout pressure on the two sets of lists (anon and file) depending on the fraction of pages on each list that were recently referenced. One thing missing is good tests for a page replacement system. With hard disks having latencies around 100,000 times higher than RAM latencies, the old "this algorithm has a 95% hit rate, while this one has an 80% hit rate" measurement strategies are no longer adequate, since either hit rate would result in absolutely atrocious performance. The lack of a good way to test page replacements systems means it is very difficult to decide whether one page replacement system is better than another. This lack of information, in turn, makes it difficult to decide on an implementation for the VM's pageout mechanism. Some of the problems (but no real solutions) have been gathered here: http://linux-mm.org/AdvancedPageReplacement -- All rights reversed. -- To unsubscribe from this list: send an email with "unsubscribe kernelnewbies" to ecartis@xxxxxxxxxxxx Please read the FAQ at http://kernelnewbies.org/FAQ
