lkml.org 
[lkml]   [2006]   [Jan]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: -mm seems significanty slower than mainline on kernbench
Date
On Thu, 12 Jan 2006 11:54 am, Peter Williams wrote:
> Peter Williams wrote:
> > Con Kolivas wrote:
> >> On Wednesday 11 January 2006 23:24, Peter Williams wrote:
> >>> Martin J. Bligh wrote:
> >>>> That seems broken to me ?
> >>>
> >>> But, yes, given that the problem goes away when the patch is removed
> >>> (which we're still waiting to see) it's broken. I think the problem is
> >>> probably due to the changed metric (i.e. biased load instead of simple
> >>> load) causing idle_balance() to fail more often (i.e. it decides to not
> >>> bother moving any tasks more often than it otherwise would) which would
> >>> explain the increased idle time being seen. This means that the fix
> >>> would be to review the criteria for deciding whether to move tasks in
> >>> idle_balance().
> >>
> >> Look back on my implementation. The problem as I saw it was that one
> >> task alone with a biased load would suddenly make a runqueue look much
> >> busier than it was supposed to so I special cased the runqueue that
> >> had precisely one task.
> >
> > OK. I'll look at that.
>
> Addressed in a separate e-mail.
>
> > But I was thinking more about the code that (in the original) handled
> > the case where the number of tasks to be moved was less than 1 but more
> > than 0 (i.e. the cases where "imbalance" would have been reduced to zero
> > when divided by SCHED_LOAD_SCALE). I think that I got that part wrong
> > and you can end up with a bias load to be moved which is less than any
> > of the bias_prio values for any queued tasks (in circumstances where the
> > original code would have rounded up to 1 and caused a move). I think
> > that the way to handle this problem is to replace 1 with "average bias
> > prio" within that logic. This would guarantee at least one task with a
> > bias_prio small enough to be moved.
> >
> > I think that this analysis is a strong argument for my original patch
> > being the cause of the problem so I'll go ahead and generate a fix. I'll
> > try to have a patch available later this morning.
>
> Attached is a patch that addresses this problem. Unlike the description
> above it does not use "average bias prio" as that solution would be very
> complicated. Instead it makes the assumption that NICE_TO_BIAS_PRIO(0)
> is a "good enough" for this purpose as this is highly likely to be the
> median bias prio and the median is probably better for this purpose than
> the average.

This is a shot in the dark. We haven't confirmed 1. there is a problem 2. that
this is the problem nor 3. that this patch will fix the problem. I say we
wait for the results of 1. If the improved smp nice handling patch ends up
being responsible then it should not be merged upstream, and then this patch
can be tested on top.

Martin I know your work move has made it not your responsibility to test
backing out this change, but are you aware of anything being done to test
this hypothesis?

Con
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2006-01-12 02:20    [from the cache]
©2003-2011 Jasper Spaans