On Wed, 4 Aug 2010 15:38:29 +0100 Mel Gorman <mel@xxxxxxxxx> wrote: > Commenting on the series "Reduce writeback from page reclaim context V6" > Andrew Morton noted; > > direct-reclaim wants to write a dirty page because that page is in the > zone which the caller wants to allocate from! Telling the flusher threads > to perform generic writeback will sometimes cause them to just gum the > disk up with pages from different zones, making it even harder/slower to > allocate a page from the zones we're interested in, no? > > On the machines used to test the series, there were relatively few zones > and only one BDI so the scenario describes is a possibility. This series is > a very early prototype series aimed at mitigating the problem. > > Patch 1 adds wakeup_flusher_threads_pages() which takes a list of pages > from page reclaim. Each inode belonging to a page on the list is marked > I_DIRTY_RECLAIM. When the flusher thread wakes, inodes with this tag are > unconditionally moved to the wb->b_io list for writing. > > Patch 2 notes that writing back inodes does not necessarily write back > pages belonging to the zone page reclaim is concerned with. In response, it > adds a zone and counter to wb_writeback_work. As pages from the target zone > are written, the zone-specific counter is updated. When the flusher thread > then checks the zone counters if a specific zone is being targeted. While > more pages may be written than necessary, the assumption is that the pages > need cleaning eventually, the inode must be relatively old to have pages at > the end of the LRU, the IO will be relatively efficient due to less random > seeks and that pages from the target zone will still be cleaned. > > Testing did not show any significant differences in terms of reducing dirty > file pages being written back but the lack of multiple BDIs and NUMA nodes in > the test rig is a problem. Maybe someone else has access to a more suitable > test rig. > > Any comment as to the suitability for such a direction? um. Might work. Isn't pretty though. But until we can demonstrate the problem or someone reports it, we probably have more important issues to be looking at ;) I think that a better approach is to try to trigger this problem as we develop and test reclaim. And if we _can't_ demonstrate it, work out why the heck not - either the code's smarter than we thought it was or the test is no good. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxxx For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>