Date: Wed, 30 Mar 2005 08:54:26 +0200
From: Ingo Molnar <mingo@elte.hu>
To: Lee Revell <rlrevell@joe-job.com>
Cc: linux-kernel@vger.kernel.org, "Paul E. McKenney" <paulmck@us.ibm.com>
Subject: Re: [patch] Real-Time Preemption, -RT-2.6.12-rc1-V0.7.41-10
Message-ID: <20050330065426.GB18417@elte.hu>
References: <20050325145908.GA7146@elte.hu> <1111790009.23430.19.camel@mindpipe> <20050325223959.GA24800@elte.hu> <1111814065.24049.21.camel@mindpipe> <20050327085814.GA23082@elte.hu> <1112159812.5598.17.camel@mindpipe>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <1112159812.5598.17.camel@mindpipe>
User-Agent: Mutt/1.4.2.1i
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1707
Lines: 35


* Lee Revell <rlrevell@joe-job.com> wrote:

> > could you run a bit with tracing disabled (in the .config) on the C3?  
> > (but wakeup timing still enabled) It may very well be tracing overhead 
> > that makes those latencies that high.  Also, we'd thus have some hard 
> > data on how much overhead tracing is in such a situation, on that CPU.
> 
> I have not left it to run overnight yet with the swappiness set to 
> 100, which triggers the biggest latencies as my entire desktop is 
> swapped out, but so far it looks like the problem was tracing 
> overhead.  With timing enabled but tracing disabled the longest 
> latency on the C3 so far is 270 usecs.
> 
> An important giveaway is that with tracing enabled the same code path 
> only triggers ~200 usec latencies on the K7 but ~2ms on the C3.  Since 
> the longest latency with PREEMPT_DESKTOP is normally more a function 
> of memory bandwidth than processor speed, and the machines differ much 
> more in the latter, this agrees with the theory that the overhead is 
> the problem.

besides cycle overhead, function tracing increases cache footprint - and 
with a CPU that has smaller caches (such as the C3) it can tip a loop 
over the edge, and can make it cache-trashing, while it would fit into 
the cache before. In such a situation the difference can be dramatic.

(on CPUs with larger caches similar artifacts can happen too, but it 
needs a 'fatter' loop, which are apparently rarer.)

	Ingo
-
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/