Date: Fri, 29 Apr 2016 03:29:38 +0800
From: Yuyang Du <yuyang.du@intel.com>
To: Peter Zijlstra <peterz@infradead.org>
Cc: mingo@kernel.org, linux-kernel@vger.kernel.org, bsegall@google.com,
        pjt@google.com, morten.rasmussen@arm.com, vincent.guittot@linaro.org,
        dietmar.eggemann@arm.com, lizefan@huawei.com, umgwanakikbuti@gmail.com
Subject: Re: [PATCH v3 4/6] sched/fair: Remove scale_load_down() for load_avg
Message-ID: <20160428192938.GC16093@intel.com>
References: <1459829551-21625-1-git-send-email-yuyang.du@intel.com>
 <1459829551-21625-5-git-send-email-yuyang.du@intel.com>
 <20160428102532.GY3430@twins.programming.kicks-ass.net>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20160428102532.GY3430@twins.programming.kicks-ass.net>
User-Agent: Mutt/1.5.21 (2010-09-15)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1547
Lines: 28

On Thu, Apr 28, 2016 at 12:25:32PM +0200, Peter Zijlstra wrote:
> On Tue, Apr 05, 2016 at 12:12:29PM +0800, Yuyang Du wrote:
> > Currently, load_avg = scale_load_down(load) * runnable%. The extra scaling
> > down of load does not make much sense, because load_avg is primarily THE
> > load and on top of that, we take runnable time into account.
> > 
> > We therefore remove scale_load_down() for load_avg. But we need to
> > carefully consider the overflow risk if load has higher range
> > (2*SCHED_FIXEDPOINT_SHIFT). The only case an overflow may occur due
> > to us is on 64bit kernel with increased load range. In that case,
> > the 64bit load_sum can afford 4251057 (=2^64/47742/88761/1024)
> > entities with the highest load (=88761*1024) always runnable on one
> > single cfs_rq, which may be an issue, but should be fine. Even if this
> > occurs at the end of day, on the condition where it occurs, the
> > load average will not be useful anyway.
> 
> I do feel we need a little more words on the actual ramification of
> overflowing here.
> 
> Yes, having 4m tasks on a single runqueue will be somewhat unlikely, but
> if it happens, then what will the user experience? How long (if ever)
> does it take for numbers to correct themselves etc..

Well, regarding the experience, this should be a stress test study.

But if the system can miraculously survive, and we end up in the scenario
that we have a ~0ULL load_sum and the rq suddently dropps to 0 load, it
would take roughly 2 seconds (=32ms*64) to converge. This time is the bound.