[lkml]   [2000]   [Oct]   [25]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
    SubjectRe: [PATCH] Re: PATCH: killing read_ahead[]

    Rik van Riel wrote:
    > On Wed, 25 Oct 2000, Jeff V. Merkey wrote:
    > > Rik van Riel wrote:
    > > > On Wed, 25 Oct 2000, Jeff V. Merkey wrote:
    > > >
    > > > > I've reviewed the patch. It's affect seems minimal and will not
    > > > > break NWFS as proposed -- it looks like, however, it will reduce
    > > > > the performance slightly of EXT2/3 with iozone for read ahead
    > > > > since the first section of the patch limits the read ahead
    > > > > window size.
    > > >
    > > > Ummm, please read it again ;)
    > > >
    > > > The patch actually /increases/ the readahead size when
    > > > we start to read a file from the beginning.
    > >
    > > But only if the file is smaller than MIN_READAHEAD * 2, which
    > > would be the case for small files (which would read the whole
    > > file anyway, which is how the page cache behaves today anyway).
    > If the file is bigger than MIN_READAHEAD * 2, we will want
    > to read in the file in multiple IOs anyway.
    > If it turns out that we read that file sequentially, then
    > the kernel will read in a LARGER CHUNK next time, if it
    > turns out that we aren't using the file sequentially, we
    > won't.
    > The point of chosing the MIN_READAHEAD * 2 cutoff is that
    > I want to avoid the N+1 problem, where we do an IO for the
    > first MIN_READAHEAD pages and have to do a separate IO for
    > the last 1 page...

    This makes sense. One issue however, for ndb with read ahead relates to
    mirroring. On NWFS,
    my read-ahead window is always (cluster size + 1). This means for
    sequential access, I will
    almost always read 64K + 1, but I do NOT allow round robin reads from
    mirrors to interleave
    4K reads between devices. I read in chunks of 64K from each mirror
    (since NetWare clusters are almost always contiguous runs of sectors in
    each cluster unless a 4K block has been hotfixed).

    The raid agent in Linux should do the same if possible, and the changes
    in the page cache and it's read ahead behavior really need to be
    interlocked with the Raid driver underneath, so when
    it reads from mirrored devices, it does not interleave 4K -> disk 1 and
    4K -> disk 2 if you know that several 4K blocks are sequential on the
    disk. NWFS does this naturally since it's cluster size layouts
    guarantee 64K of continuous 4K pages on the disk. This explains why on
    2.4 NWFS mirroring with read ahead enabled is several orders faster than
    running without mirroring.

    I've seen some performance issues with Raid on Linux with mirroring
    because the driver is interleaving 4K reads between devices, rather than
    taking them in big chunks from each spindle.



    > regards,
    > Rik
    > --
    > "What you're running that piece of shit Gnome?!?!"
    > -- Miguel de Icaza, UKUUG 2000
    To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
    the body of a message to
    Please read the FAQ at

     \ /
      Last update: 2005-03-22 12:41    [W:0.024 / U:10.484 seconds]
    ©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site