On Mon, 10 Dec 2001, Mike Kravetz wrote:
> On Sun, Dec 09, 2001 at 05:38:42PM -0800, Davide Libenzi wrote:
> > Coming at the pipe example, let's take Larry's lat_ctx ( lmbench ).
> > This is bouncing data through pipes using I/O bound tasks, and running
> > vmstat together with a lat_ctx 32 32 ... ( long list ), you'll see the run
> > queue length barley reach 3 ( with 32 bouncing tasks ).
> > It barely reaches 5 with 64 bouncing tasks.
> This may show my ignorance, but ... Why would one expect much
> more than 2 runnable tasks as a result of a running lat_ctx?
> This benchmark simply passes a token around a ring of tasks.
> One task awakens the next, then goes to sleep. The only time
> you have more than one runnable task is during the times when
> the token is passed between tasks. In these transition times
> I would rarely expect more than 2 tasks on the runqueue no
> matter how many bouncing tasks you have.
> We created a benchmark similar to lat_ctx that would allow you
> to control how many runnable tasks there are in the system.
> Look for 'Reflex' benchmark at:
> You can think of this as a controlled way of running multiple
> copies of lat_ctx in parallel.
Mike, this is not to criticize lat_ctx ( i'll do it later :) ) but is
started from an example that Alan made about processes bouncing data through pipes.
Talking about lat_ctx, it's a benchmark that you've to use 1) not to plot
latencies by having runqueue lengths on the x axis 2) only on UP systems (
take a look at the process distribution on SMP ).
I usually prefer to have a real measure of the scheduler latency and
having a real runqueue length to put on the x axis.
I cannot have this with lat_ctx because :
1) you'll never have a given rq len
2) numbers are bogus
If you've to show how the latency varies with the runqueue length you've
to have a benchmark that during the test shows that length.
That's why i prefer to use a cycle counter + the cpuhog program.
In that way i do not need to switch like a crazy ( and you're actually
going to measure warm cache performances ) to get the measure.
I can simply sample my system while it's switching "naturally" with my
system switch-load plus the cpuhog load on the runqueue.
So lat_ctx is basically a 1) simple 2) warm cache 3) UP 4) not rqlen
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to email@example.com
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Sat Dec 15 2001 - 21:00:19 EST