Re: [patch] Real-Time Preemption, -RT-2.6.12-rc1-V0.7.41-10

[Ingo Molnar]

* Lee Revell <rlrevell@xxxxxxxxxxx> 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@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/