Date: Tue, 6 Dec 2016 15:34:29 +0100
From: Frederic Weisbecker <fweisbec@gmail.com>
To: Paul Mackerras <paulus@ozlabs.org>
Cc: LKML <linux-kernel@vger.kernel.org>, Tony Luck <tony.luck@intel.com>,
        Wanpeng Li <wanpeng.li@hotmail.com>,
        Peter Zijlstra <peterz@infradead.org>,
        Michael Ellerman <mpe@ellerman.id.au>,
        Heiko Carstens <heiko.carstens@de.ibm.com>,
        Benjamin Herrenschmidt <benh@kernel.crashing.org>,
        Thomas Gleixner <tglx@linutronix.de>, Ingo Molnar <mingo@kernel.org>,
        Fenghua Yu <fenghua.yu@intel.com>, Rik van Riel <riel@redhat.com>,
        Martin Schwidefsky <schwidefsky@de.ibm.com>,
        Stanislaw Gruszka <sgruszka@redhat.com>
Subject: Re: [PATCH 00/10] vtime: Delay cputime accounting to tick
Message-ID: <20161206143428.GA2507@lerouge>
References: <1480991543-6557-1-git-send-email-fweisbec@gmail.com>
 <20161206042055.GB9068@fergus.ozlabs.ibm.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20161206042055.GB9068@fergus.ozlabs.ibm.com>
User-Agent: Mutt/1.5.24 (2015-08-30)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 2142
Lines: 63

On Tue, Dec 06, 2016 at 03:20:55PM +1100, Paul Mackerras wrote:
> On Tue, Dec 06, 2016 at 03:32:13AM +0100, Frederic Weisbecker wrote:
> > This follows up Martin Schwidefsky's patch which propose to delay
> > cputime accounting to the tick in order to minimize the calls to
> > account_system_time() and alikes as these functions can carry quite some
> > overhead:
> > 
> > 	http://lkml.kernel.org/r/20161121111728.13a0a3db@mschwide
> > 
> > The set includes Martin's patch, rebased on top of tip:sched/core and
> > latest s390 changes, and extends it to the other implementations of
> > CONFIG_VIRT_CPU_ACCOUNTING_NATIVE (powerpc and ia64) along with a few
> > core changes to adapt the whole.
> > 
> > Only built-tested though as I don't have access to any of these archs.
> 
> The patches look reasonable at a quick look.  I assume that to test
> them, we would want to run a guest in an overcommitted system, so as
> to get some steal time.  Do you have any more specific suggestions as
> to what to run as a test?  Just run some benchmark and see if the
> user/system/irq times look reasonable?  Or do you have something more
> quantitative?

So I guess we want to test both correctness and performance.

To check correctness I use two little programs, one that does a userspace
loop:

int main(int argc, char **argv)
{
	while (1);
	return 0;
}


And another that does a kernelspace loop. The latter
is not 100% kernel loop but spends most of its time in
kernel mode.

int main(int argc, char **argv)
{
	void *addr = sbrk(0);

	while (1) {
		brk(addr + 4096);
		brk(addr);
	}
	return 0;
}

Testing idle time just consist in checking the difference between two
cat /proc/stat in a given timelapse for an idle CPU.

For irqs it gets harder. There you just need to check if the numbers are
reasonable.

Now in order to measure performance, I think you need a workload that either
does a lot of guest/host switch or does a lot of IRQs. Maybe just something
that involves networking. Then comparing stime, hardirq and softirq should
show some better nummbers. In order to increase the effect, you can set a very
low HZ value (100?).

Thanks.