lkml.org 
[lkml]   [2014]   [Oct]   [26]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
SubjectRe: bisected: futex regression >= 3.14 - was - Slowdown due to threads bouncing between HT cores
From
Date
On Sun, 2014-10-26 at 14:16 +0100, Mike Galbraith wrote: 
> On Sun, 2014-10-26 at 11:39 +0100, Steinar H. Gunderson wrote:
> > On Fri, Oct 24, 2014 at 06:38:41PM +0200, Mike Galbraith wrote:
> > >>> Whew, good, futex.c is hard. Heads up chess guys <punt>.
> > >> I wonder whether the barrier fix which got into 3.17 late fixes that
> > >> issue as well.
> > > Yes, it did.
> >
> > This is only about the lockup, right, not that the threads bounce around a
> > lot and make things slower?
>
> Yes. I couldn't reproduce your results, on my 28 core+ht box, there was
> little difference between pinned/unpinned. Maybe the box was too quite,
> not enough random activity to motivate select_idle_sibling().

Can you try the below?

---
kernel/sched/fair.c | 6 ++++++
1 file changed, 6 insertions(+)

--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -4477,6 +4477,12 @@ static int select_idle_sibling(struct ta
return i;

/*
+ * Don't bounce hefty tasks about, they may be hitting contention.
+ */
+ if (p->se.load.weight - (p->se.load.weight/4) < p->se.avg.load_avg_contrib)
+ return target;
+
+ /*
* Otherwise, iterate the domains and find an elegible idle cpu.
*/
sd = rcu_dereference(per_cpu(sd_llc, target));



\
 
 \ /
  Last update: 2014-10-26 15:41    [W:0.045 / U:0.228 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site