lkml.org 
[lkml]   [2008]   [Dec]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    Patch in this message
    /
    Date
    From
    SubjectRe: [RFC PATCH v5 2/7] sched: favour lower logical cpu number for sched_mc balance
    * Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com> [2008-12-11 23:12:48]:

    > Just in case two groups have identical load, prefer to move load to lower
    > logical cpu number rather than the present logic of moving to higher logical
    > number.
    >
    > find_busiest_group() tries to look for a group_leader that has spare capacity
    > to take more tasks and freeup an appropriate least loaded group. Just in case
    > there is a tie and the load is equal, then the group with higher logical number
    > is favoured. This conflicts with user space irqbalance daemon that will move
    > interrupts to lower logical number if the system utilisation is very low.
    >

    This patch will work well with irqbalance only when irqbalance decides
    to switch to power mode and if the interrupt rate is high and
    irqbalance is in performance mode and sched_mc > 1, what is the impact
    of this patch?

    > Signed-off-by: Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com>
    > ---
    >
    > kernel/sched.c | 4 ++--
    > 1 files changed, 2 insertions(+), 2 deletions(-)
    >
    > diff --git a/kernel/sched.c b/kernel/sched.c
    > index 322cd2a..6bea99b 100644
    > --- a/kernel/sched.c
    > +++ b/kernel/sched.c
    > @@ -3264,7 +3264,7 @@ find_busiest_group(struct sched_domain *sd, int this_cpu,
    > */
    > if ((sum_nr_running < min_nr_running) ||
    > (sum_nr_running == min_nr_running &&
    > - first_cpu(group->cpumask) <
    > + first_cpu(group->cpumask) >
    > first_cpu(group_min->cpumask))) {

    The first_cpu logic worries me a bit. This has existed for a while
    already, but with the topology I see on my system, I find the cpu
    numbers interleaved on my system (0,2,4 and 6) belong to one core and
    odd numbers to the other.

    So for a topology like (assume dual core, dual socket)

    0-3
    / \
    0-1 2-3
    / \ / \
    0 1 2 3

    If group_min is the domain with (2-3) and we are looking at
    group(0-1). first_cpu of (0-1) is 0 and (2-3) is 2, how does changing
    "<" to ">" help push the tasks to the lower ordered group? In the case
    described above group_min continues to be (2-3).

    Shouldn't the check be if (first_cpu(group->cpumask) <=
    first_cpu(group_min->cpumask)?

    > group_min = group;
    > min_nr_running = sum_nr_running;
    > @@ -3280,7 +3280,7 @@ find_busiest_group(struct sched_domain *sd, int this_cpu,
    > if (sum_nr_running <= group_capacity - 1) {
    > if (sum_nr_running > leader_nr_running ||
    > (sum_nr_running == leader_nr_running &&
    > - first_cpu(group->cpumask) >
    > + first_cpu(group->cpumask) <
    > first_cpu(group_leader->cpumask))) {
    > group_leader = group;
    > leader_nr_running = sum_nr_running;
    >
    >

    All these changes are good, I would like to see additional statistics
    that show how many decisions were taken due to new power aware
    balancing logic, so that I spot the bad and corner cases based on the
    statistics I see.


    --
    Balbir


    \
     
     \ /
      Last update: 2008-12-15 07:15    [W:0.028 / U:0.268 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site