[lkml]   [2010]   [Oct]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
    SubjectRe: [RFC/RFT PATCH] sched: automated per tty task groups
    On Tue, 2010-10-19 at 08:28 -0700, Linus Torvalds wrote:
    > On Tue, Oct 19, 2010 at 4:29 AM, Mike Galbraith <> wrote:

    > So I think the patch looks pretty good, and the numbers seem to look
    > just stunningly so, but I'd like to name the feature more along the
    > lines of "automatic process group scheduling" rather than about tty's
    > per se.

    Oh, absolutely, that's what it's all about really. What I'd _like_ is
    to get per process group scheduling working on the cheap..ish. Your
    idea of tty cgoups looked much simpler though, so I figured that would
    be a great place to start. It turned out to be much simpler than I
    thought it would be, which is encouraging, and it works well in testing
    (so far that is).

    > And you actually did that for the Kconfig option, which makes me quite happy.

    (Ingo's input.. spot on)

    > The one other thing I do wonder about is how noticeable the group
    > scheduling overhead is.

    Very noticeable, cgroups is far from free. It would make no sense for a
    performance freak to even think about it. I don't run cgroup enabled
    kernels usually, and generally strip to the bone because I favor
    throughput very very heavily, but when I look at the desktop under load,
    the cost/performance trade-off ~seems to work out.

    > If people compare with a non-CGROUP_SCHED
    > kernel, will a desktop-optimized kernel suddenly have horrible pipe
    > latency due to much higher scheduling cost? Right now that whole
    > feature is hidden by EXPERIMENTAL, I don't know how much it hurts, and
    > I never timed it when I tried it out long ago..

    The scheduling cost is quite high. But realistically, the cost of a
    distro kernel with full featured network stack is (much) higher. I
    seriously doubt the cost of cgroups would be noticed by the typical
    _desktop_ user. Overall latencies for any switchy microbenchmark will
    certainly be considerably higher with the feature enabled.


     \ /
      Last update: 2010-10-19 20:15    [W:0.109 / U:15.372 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site