[lkml]   [2009]   [Apr]   [7]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
    SubjectRE: [PATCH 0/8][RFC] IO latency/throughput fixes
    The (small) set of people that rely on "ordered" understand 
    the problem, as long as they are aware of the change (no, I
    don't think reading through all changelogs from their old
    kernel to the new one is a realistic option).

    So a config option should be good enough to get them to notice
    the change (I assume a missing default will force them to
    choose an option), and therefore explicitly add the -o ordered
    option to their scripts.

    On the other hand a run-time tunable has no real point.

    > -----Original Message-----
    > From: Linus Torvalds []
    > Sent: Monday, April 06, 2009 10:02 PM
    > To: Trenton D. Adams
    > Cc: Chris Mason; Theodore Tso; Hua Zhong; Jens Axboe; Linux Kernel
    > Mailing List
    > Subject: Re: [PATCH 0/8][RFC] IO latency/throughput fixes
    > On Mon, 6 Apr 2009, Trenton D. Adams wrote:
    > >
    > > What about a procfs setting instead? Is there a policy about why
    > > something should be in procfs or /sys, or as a kernel config option?
    > > That's basically as small as the patch you just made, right?
    > I'm never really against making things dynamically tunable, but this
    > already was, and that wasn't really the issue.
    > Sure, you can just re-mount your filesystem with different options.
    > That's
    > what I did while testing - my /home is on a drive of its own, so I
    > would
    > just log out and as root unmount and re-mount with
    > data=ordered/writeback,
    > and log in and test again.
    > So dynamic tuning is good. But at the same time, having a tuning option
    > is
    > _never_ an excuse for not getting the default right in the first place.
    > It's just a cop-out to say "hey, the default may be wrong for you, but
    > you
    > can always tune it locally with XYZ".
    > The thing is, almost nobody does that. Partly because it needs effort
    > and
    > knowledge, partly because after a few years the number of tuning knobs
    > are
    > in the hundreds for every little thing.
    > So instead, leave the tuning for the _really_ odd cases (if you use
    > your
    > machine as an IP router, you hopefully know enough to tune it if you
    > really care). Not for random general-purpose "use for whatever" kind of
    > thing.
    > > I'm just thinking that something like this, where people want one
    > > thing or the other, but may not know it when they install Linux,
    > might
    > > like to change it realtime. Especially if they are a Linux newbie,
    > > and don't know how to compile their own kernel. Or don't have time
    > to
    > > maintain their own kernel installs.
    > Oh absolutely. I'm not expecting people to compile their own kernels.
    > I'm
    > expecting that within a few months, most modern distributions will have
    > (almost by mistake) gotten a new set of saner defaults, and anybody who
    > keeps their machine up-to-date will see a smoother experience without
    > ever
    > even realizing _why_.
    > Linus

     \ /
      Last update: 2009-04-07 07:27    [W:0.025 / U:4.444 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site