lkml.org 
[lkml]   [2011]   [Apr]   [22]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: [PATCH 5/6] writeback: try more writeback as long as something was written
    On Fri 22-04-11 10:32:26, Wu Fengguang wrote:
    > On Fri, Apr 22, 2011 at 12:41:54AM +0800, Jan Kara wrote:
    > > On Thu 21-04-11 14:05:56, Wu Fengguang wrote:
    > > > On Thu, Apr 21, 2011 at 12:39:40PM +0800, Christoph Hellwig wrote:
    > > > > On Thu, Apr 21, 2011 at 11:33:25AM +0800, Wu Fengguang wrote:
    > > > > > I collected the writeback_single_inode() traces (patch attached for
    > > > > > your reference) each for several test runs, and find much more
    > > > > > I_DIRTY_PAGES after patchset. Dave, do you know why there are so many
    > > > > > I_DIRTY_PAGES (or radix tag) remained after the XFS ->writepages() call,
    > > > > > even for small files?
    > > > >
    > > > > What is your defintion of a small file? As soon as it has multiple
    > > > > extents or holes there's absolutely no way to clean it with a single
    > > > > writepage call.
    > > >
    > > > It's writing a kernel source tree to XFS. You can find in the below
    > > > trace that it often leaves more dirty pages behind (indicated by the
    > > > I_DIRTY_PAGES flag) after writing as less as 1 page (indicated by the
    > > > wrote=1 field).
    > > As Dave said, it's probably just a race since XFS redirties the inode on
    > > IO completion. So I think the inodes are just small so they have only a few
    > > dirty pages so you don't have much to write and they are written and
    > > redirtied before you check the I_DIRTY flags. You could use radix tree
    > > dirty tag to verify whether there are really dirty pages or not...
    >
    > Yeah, Dave and Christoph root caused it in the other email -- XFS sets
    > I_DIRTY which accidentally sets I_DIRTY_PAGES. We can safely bet there
    > are no real dirty pages -- otherwise it would have turned up as
    > performance regressions.
    Yes, but then the question what we actually do better is still open,
    right? :) I'm really curious what it could be because especially in your
    copy-kernel case I should not make much different - maybe except if we
    occasionally managed to block on PageLock behind the writing thread and now
    we don't because we queue the inode later but I find that highly unlikely.

    > > BTW a quick check of kernel tree shows the following distribution of
    > > sizes (in KB):
    > > Count KB Cumulative Percent
    > > 257 0 0.9%
    > > 13309 4 45%
    > > 5553 8 63%
    > > 2997 12 73%
    > > 1879 16 80%
    > > 1275 20 83%
    > > 987 24 87%
    > > 685 28 89%
    > > 540 32 91%
    > > 387 36 ...
    > > 309 40
    > > 264 44
    > > 249 48
    > > 170 52
    > > 143 56
    > > 144 60
    > > 132 64
    > > 100 68
    > > ...
    > > Total 30155
    > >
    > > And the distribution of your 'wrote=xxx' roughly corresponds to this...
    >
    > Nice numbers! How do you manage to account them? :)
    Easy shell command (and I handcomputed the percentages because I was lazy
    to write a script for that):
    find . -type f -name "*.[ch]" -exec du {} \; | cut -d ' ' -f 1 |
    sort -n | uniq -c

    Honza
    --
    Jan Kara <jack@suse.cz>
    SUSE Labs, CR


    \
     
     \ /
      Last update: 2011-04-22 23:25    [W:3.035 / U:0.008 seconds]
    ©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site