lkml.org 
[lkml]   [2006]   [Jun]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: [RFC] [PATCH 5/8] inode-diet: Eliminate i_blksize and use a per-superblock default
    On 6/19/06, Theodore Tso <tytso@mit.edu> wrote:
    > On Mon, Jun 19, 2006 at 09:16:51AM -0700, Joel Becker wrote:
    > > On Mon, Jun 19, 2006 at 04:58:21PM +0100, Christoph Hellwig wrote:
    > > > Blease don't add a field to the superblock for the optimal I/O size.
    > > > Any filesystem that wants to override it can do so directly in ->getattr.
    > >
    > > I don't disagree with you, but the idea of everyone implementing
    > > ->getattr where they just let it work before scares me. It's a ton of
    > > cut-n-paste error waiting to happen. Especially if we make something
    > > stale.
    > > Perhaps add generic_fillattr_blksize()?
    >
    > Well, as far as I know the only filesystems today that would need to
    > do something different are xfs, ocfs2, and reiserfs, and IMHO only the
    > first two have any kind of justification for doing it. Part of the
    > problem is what st_blksize actually means was never well-defined; it
    > was never in POSIX, and in SuSv3 all that is stated is, "A file
    > system-specific preferred I/O block size for this object." This is
    > why Reiserfs got away with specifying 128 megs (I assume it helped on
    > some benchmark), and why being ill-defined, using such a large value
    > might cause some applications (like /bin/cp) to core dump.
    >
    > Given that most filesystems use the generic page cache read/write
    > functions, using PAGE_CACHE_SIZE as the default seems to make a huge
    > amount of sense. I really wonder how useful setting st_blksize really
    > is, actually, at least in the real-world, as opposed to just for
    > benchmarks.

    I assume "128 megs" is a typo, it's 128k, of course. And certainly it
    would have helped speed things up, not just benchmarks, because for
    modern disks, doing 128k vs 4k takes like 25% more time. Wu's
    adaptive readahead patches might make this outdated, though, since
    they now support "read 10 pages, seek, read 10 pages, seek, etc" type
    workloads.

    NATE
    -
    To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
    the body of a message to majordomo@vger.kernel.org
    More majordomo info at http://vger.kernel.org/majordomo-info.html
    Please read the FAQ at http://www.tux.org/lkml/

    \
     
     \ /
      Last update: 2006-06-21 21:43    [W:0.029 / U:0.320 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site