Message-ID: <1338820234.7356.250.camel@marge.simpson.net>
Subject: Re: [PATCH] sched: balance_cpu to consider other cpus in its group
 as target of (pinned) task migration
From: Mike Galbraith <efault@gmx.de>
To: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>,
        Prashanth Nageshappa <prashanth@linux.vnet.ibm.com>, mingo@kernel.org,
        LKML <linux-kernel@vger.kernel.org>, roland@kernel.org,
        Ingo Molnar <mingo@elte.hu>
Date: Mon, 04 Jun 2012 16:30:34 +0200
In-Reply-To: <20120604130740.GC25126@linux.vnet.ibm.com>
References: <4FCC4E3B.4090209@linux.vnet.ibm.com>
	 <1338801907.7356.163.camel@marge.simpson.net>
	 <1338810796.28282.32.camel@twins>
	 <1338814063.7356.192.camel@marge.simpson.net>
	 <20120604130740.GC25126@linux.vnet.ibm.com>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: 7bit
Mime-Version: 1.0
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1808
Lines: 35

On Mon, 2012-06-04 at 18:37 +0530, Srivatsa Vaddagiri wrote: 
> * Mike Galbraith <efault@gmx.de> [2012-06-04 14:47:43]:
> 
> > You need a good reason to run RT, and being able to starve others to
> > death ain't it, so I don't see a good reason to care about the 95% case
> > enough to fiddle with load balancing to accommodate the oddball case.
> 
> While starvation of SCHED_OTHER task was an extreme case, the point
> remains that SCHED_OTHER tasks are better served by moving them away
> from cpus running rt tasks that are partially cpu intensive. While the
> current code has the nuts and bolts to recognize this situation
> (scale_rt_power), it fails to effect SCHED_OTHER task movement because of how
> one cpu from a sched_group is designated to pull tasks on behalf of its
> siblings and that chosen balance_cpu may not be in the task's cpus_allowed mask
> (but the task can run on one or more of its sibling cpus).

Yeah, this is true, it is a latency source and a fairness violation.
Slow path balance consideration does make some sense to me.

But, if you have an RT requirement, you can't afford to mix unknown
entities, nor over-commit etc.  A realtime application will assign all
resources, so the load balancer becomes essentially unemployed.  No?
Non critical worker-bees may be allowed to bounce around in say a
cpuset, but none of the CPUs which do critical work will ever be
over-committed, else application just lost the war.  In that regard,
twiddling the load balancer to accommodate strange sounding case still
seems wrong to me.

-Mike

--
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/