[lkml]   [2011]   [Jun]   [29]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
    Patch in this message
    SubjectRe: power increase issue on light load
    On Tue, Jun 28, 2011 at 7:30 PM, Nikhil Rao <> wrote:
    > Looking at the schedstat data Alex posted:
    > - Distribution of load balances across cores looks about the same.
    > - Load balancer does more idle balances on 3.0-rc4 as compared to
    > 2.6.39 on SMT and NUMA domains. Busy and newidle balances are a mixed
    > bag.
    > - I see far fewer affine wakeups on 3.0-rc4 as compared to 2.6.39.
    > About half as many affine wakeups on SMT and about a quarter as many
    > on NUMA.
    > I'm investigating the impact of the load resolution patchset on
    > effective load and wake affine calculations. This seems to be the most
    > obvious difference from the schedstat data.

    I went through the math in effective load and wake affine and I think
    it should be OK. There are a couple of corner cases where increasing
    sched load resolution can change the result of wake affine -- I've
    listed them below. However, I not convinced you are hitting these
    cases often enough to make a noticeable difference. I'm looking into
    the other LB paths...

    - One corner case is because of rounding error in the shares update
    path. Let's say the shares update logic assigned weight A to a sched
    entity in the case with scaled resolution, and it assigned weight B
    without scaling weights. Now, we expect A/1024 = B, but this is not
    always the case because of rounding error. The difference between (A
    and B*1024) gets amplified in wake_affine() since it multiplies
    (weight+effective load) with imbalance pct and cpu power -- we
    effectively scale this up by 5 orders of magnitude. In cases where
    prev_eff_load and this_eff_load are pretty close, this difference can
    result in a different result in wake_affine().

    - There's a corner case in effective_load(), where if a task wakes up
    on an empty cfs_rq, you could hit the clamp in effective_load (i.e. <
    MIN_SHARES) which can affect prev_eff_load (you get a lower number --
    making it less likely to do an affine wakeup). I think this patch
    (against 3.0-rc4) will address that issue -- can you please give this
    a try?

    diff --git a/kernel/sched_fair.c b/kernel/sched_fair.c
    index 433491c..6fcfbfc 100644
    --- a/kernel/sched_fair.c
    +++ b/kernel/sched_fair.c
    @@ -1442,8 +1442,8 @@ static long effective_load(struct task_group
    *tg, int cpu, long wl, long wg)
    wl = tg->shares;

    /* zero point is MIN_SHARES */
    - if (wl < MIN_SHARES)
    - wl = MIN_SHARES;
    + if (wl < scale_load(MIN_SHARES))
    + wl = scale_load(MIN_SHARES);
    wl -= se->load.weight;
    wg = 0;

     \ /
      Last update: 2011-06-30 02:11    [W:0.022 / U:6.044 seconds]
    ©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site