Re: [2.4.17/18pre] VM and swap - it's really unusable

yodaiken@fsmlabs.com
Mon, 21 Jan 2002 16:56:59 -0700


On Mon, Jan 21, 2002 at 05:01:45PM -0500, Robert Love wrote:
> On Mon, 2002-01-21 at 16:49, yodaiken@fsmlabs.com wrote:
>
> > > (average of 4 runs of `dbench 16')
> > > 2.5.3-pre1: 25.7608 MB/s
> > > 2.5.3-pre1-preempt: 32.341 MB/s
> > >
> > > (old, average of 4 runs of `dbench 16')
> > > 2.5.2-pre11: 24.5364 MB/s
> > > 2.5.2-pre11-preempt: 27.5192 MB/s
>
> > Robert, with all due respect, my tests of dbench show such high
> > variation that 4 miserable runs prove exactly nothing.
>
> Well you asked for dbench. Would you prefer 10 runs each? There were,

50. And I'd like standard deviation as well as average, best, worst.

And then I'd like to see the same test done with a RT task running in
the background - since I assume preempt has as its main purpose
to enable SCHED_FIFO.

> I guess the point is, everyone argues preemption is detrimental to
> throughput. I'm not going to argue that we aren't adding complexity,

I have not seen that argued - certainly I have not argued it myself.
My argument is:
It makes the kernel _much_ more complex
It has known costs e.g. by making the lockless
per-processor caching more difficult if not impossible
It seems to lead to a requirement for inheritance
It has no demonstrated benefits.

> > Did these even come on the same filesystem?
>
> Yes, why would you suspect otherwise?
>

Because the prior cited benchmark was a kernel compile on different trees.
Was the filesystem unchanged between runs?
I'm not suggesting you are cheating, it's easy to overlook something
critical when there are so many variables.

-- 
---------------------------------------------------------
Victor Yodaiken 
Finite State Machine Labs: The RTLinux Company.
 www.fsmlabs.com  www.rtlinux.com

- 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/