On Tue, Sep 22, 2009 at 11:59:41PM +0800, Chris Mason wrote: > On Tue, Sep 22, 2009 at 09:18:32PM +0800, Wu Fengguang wrote: > > On Tue, Sep 22, 2009 at 07:30:55PM +0800, Chris Mason wrote: > > [ using a very large MAX_WRITEBACK_PAGES ] > > > > > > I'm starting to rethink the 128MB MAX_WRITEBACK_PAGES. 128MB is the > > > > > right answer for the flusher thread on sequential IO, but definitely not > > > > > on random IO. We don't want the flusher to get bogged down on random > > > > > writeback and start ignoring every other file. > > > > > > > > Hmm, I'd think a larger MAX_WRITEBACK_PAGES shall never increase the > > > > writeback randomness. > > > > > > It doesn't increase the randomness, but if we have a file full of > > > buffered random IO (say from bdb or rpm), the 128MB max will mean that > > > one file dominates the flusher thread writeback completely. > > > > What if we add a bdi->max_segments quota? A segment is a continuous > > run of dirty pages in the inode address space. SSD or fast RAID could > > set it to a large enough value. > > I'd rather play with timeslice ideas first ;) But, don't let me stop > you from trying interesting things. OK. > > > > > > > > > > > My btrfs performance branch has long had a change to bump the > > > > > nr_to_write up based on the size of the delayed allocation that we're > > > > > doing. It helped, but not as much as I really expected it too, and a > > > > > similar patch from Christoph for XFS was good but not great. > > > > > > > > > > It turns out the problem is in write_cache_pages. It processes a whole > > > > > pagevec at a time, something like this: > > > > > > > > > > while(!done) { > > > > > for each page in the pagegvec { > > > > > writepage() > > > > > if (wbc->nr_to_write <= 0) > > > > > done = 1; > > > > > } > > > > > } > > > > > > > > > > If the filesystem decides to bump nr_to_write to cover a whole > > > > > extent (or a max reasonable size), the new value of nr_to_write may > > > > > be ignored if nr_to_write had already gone done to zero. > > > > > > > > > > I fixed btrfs to recheck nr_to_write every time, and the results are > > > > > much smoother. This is what it looks like to write out all the .o files > > > > > in the kernel. > > > > > > > > > > http://oss.oracle.com/~mason/seekwatcher/btrfs-nr-to-write.png > > > > > > > > > > In this graph, Btrfs is writing the full extent or 8192 pages, whichever > > > > > is smaller. The write_cache_pages change is here, but it is local to > > > > > the btrfs copy of write_cache_pages: > > > > > > > > > > http://git.kernel.org/?p=linux/kernel/git/mason/btrfs-unstable.git;a=commit;h=f85d7d6c8f2ad4a86a1f4f4e3791f36dede2fa76 > > > > > > > > It seems you tried to an upper limit of 32-64MB: > > > > > > > > + if (wbc->nr_to_write < delalloc_to_write) { > > > > + int thresh = 8192; > > > > + > > > > + if (delalloc_to_write < thresh * 2) > > > > + thresh = delalloc_to_write; > > > > + wbc->nr_to_write = min_t(u64, delalloc_to_write, > > > > + thresh); > > > > + } > > > > > > > > However it is possible that btrfs bumps up nr_to_write for each inode, > > > > so that the accumulated bump ups are too large to be acceptable for > > > > balance_dirty_pages(). > > > > > > We bump up to a limit of 64MB more than the original nr_to_write. This > > > is because when we do bump we know we'll write the whole amount, and > > > then write_cache_pages will end. > > > > Imagine this scenario. There are inodes A, B, C, ... > > > > A) delalloc_to_write=3000 but only 1000 pages dirty. > > The part that isn't clear from the code you're reading is that if > delalloc_to_write is 3000, then there must be 3000 pages dirty. The > count of delalloc bytes to go down always reflects IO that must be done. > > So, once my writepage call bumps nr_to_write, that IO will happen. The > only exception is if someone else jumps in and writes the pages, which > won't happen unless there is synchronous writeback. Ah thanks for the clarification. > > > > Yes a more general solution would help. I'd like to propose one which > > > > works in the other way round. In brief, > > > > (1) the VFS give a large enough per-file writeback quota to btrfs; > > > > (2) btrfs tells VFS "here is a (seek) boundary, stop voluntarily", > > > > before exhausting the quota and be force stopped. > > > > > > > > There will be two limits (the second one is new): > > > > > > > > - total nr to write in one wb_writeback invocation > > > > - _max_ nr to write per file (before switching to sync the next inode) > > > > > > > > The per-invocation limit is useful for balance_dirty_pages(). > > > > The per-file number can be accumulated across successive wb_writeback > > > > invocations and thus can be much larger (eg. 128MB) than the legacy > > > > per-invocation number. > > > > > > > > The file system will only see the per-file numbers. The "max" means > > > > if btrfs find the current page to be the last page in the extent, > > > > it could indicate this fact to VFS by setting wbc->would_seek=1. The > > > > VFS will then switch to write the next inode. > > > > > > > > The benefit of early voluntarily yield is, it reduced the possibility > > > > to be force stopped half way in an extent. When next time VFS returns > > > > to sync this inode, it will again be honored the full 128MB quota, > > > > which should be enough to cover a big fresh extent. > > > > > > This is interesting, but it gets into a problem with defining what a > > > seek is. On some hardware they are very fast and don't hurt at all. It > > > might be more interesting to make timeslices. > > > > We could have quotas for max pages, page segments and submission time. > > Will they be good enough? The first two quotas could be made per-bdi > > to reflect hardware capabilities. > > The reason I prefer the timeslice idea is that we don't need the > hardware to tell us how fast it is. We just write for a while and move > on. That makes sense. Note that the triple (pages, page segments, submission time) can somehow adapt to hardware capabilities (and at least won't hurt fast arrays). - max pages are set to large enough number for big arrays - max page segments could be based on the existing blk_queue_nonrot() - submission time = 1s, which is mainly a safeguard for slow devices (ie. usb stick), to prevent one single inode from taking too much time. This time limit has little performance impacts. Possible merits are - these parameters are concrete ones and easy to handle - it's natural to implement related logics in the VFS level - file systems can do nothing to get most benefits Also the (now necessary) per-invocation limit could be somehow eliminated when balance_dirty_pages() does not do IO itself. Thanks, Fengguang -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html