Re: [PATCH] x86/xen: Add "xen_timer_slop" command line option

From: Ryan Thibodeaux <thibodux@gmail.com>
To: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Cc: xen-devel@lists.xenproject.org, linux-kernel@vger.kernel.org,
	oleksandr_andrushchenko@epam.com, tglx@linutronix.de,
	jgross@suse.com, ryan.thibodeaux@starlab.io
Subject: Re: [PATCH] x86/xen: Add "xen_timer_slop" command line option
Date: Sat, 23 Mar 2019 08:00:52 -0400	[thread overview]
Message-ID: <20190323120052.GB11100@centos-dev.localdomain> (raw)
In-Reply-To: <abf65a68-9463-0f0e-1f51-6eefd3cacc88@oracle.com>

On Fri, Mar 22, 2019 at 06:10:16PM -0400, Boris Ostrovsky wrote:
> On 3/22/19 2:29 PM, thibodux@gmail.com wrote:
> > From: Ryan Thibodeaux <ryan.thibodeaux@starlab.io>
> >
> > Add a new command-line option "xen_timer_slop=<INT>" that sets the
> > minimum delta of virtual Xen timers. This commit does not change the
> > default timer slop value for virtual Xen timers.
> >
> > Lowering the timer slop value should improve the accuracy of virtual
> > timers (e.g., better process dispatch latency), but it will likely
> > increase the number of virtual timer interrupts (relative to the
> > original slop setting).
> >
> > The original timer slop value has not changed since the introduction
> > of the Xen-aware Linux kernel code. This commit provides users an
> > opportunity to tune timer performance given the refinements to
> > hardware and the Xen event channel processing. It also mirrors
> > a feature in the Xen hypervisor - the "timer_slop" Xen command line
> > option.
> 
> Is there any data that shows effects of using this new parameter?
> 
> -boris
> 

For our own testing using "cyclictest" from the rt-tests project,
lowering the timer slop helped produce the best test runs, especially
in terms of maximum process dispatch latency (PDL).

Here is the output from one such test that ran overnight. The Xen
timer slop in this case was 10000 or 10 microseconds.

...
[root@slop1 ~]# cset shield -c 3
[root@slop1 ~]# echo ; date ; echo ; \
./rt-tests-1.3/cyclictest -p95 -a3 -t1 -m; echo ; date

Thu Mar 14 19:45:36 UTC 2019

# /dev/cpu_dma_latency set to 0us
policy: fifo: loadavg: 0.00 0.02 0.00 1/91 4260
T: 0 ( 3212) P:95 I:1000 C:57077313 Min: 27 Act: 44 Avg: 43 Max: 145
^C
Fri Mar 15 11:36:53 UTC 2019
...

This test system was configured to use a TSC clocksource, disabled
C states, and lowered the timer slop. I am not claiming the timer
slop change was solely responsible for the best results. In other
testing with the default timer slop setting of 100000 (100
microseconds), the average PDL would run slightly higher, but the
spikes were much higher and more in number, often near the 1000s
and happening multiple times per 10 minutes of testing.

- Ryan