Re: [Xen-devel] Re: [Xen-users] About profiling xen

[Marco Tizzoni <marco.tizzoni@xxxxxxxxx>]

On Thu, Oct 1, 2009 at 11:16 PM, Jeremy Fitzhardinge <jeremy@xxxxxxxx> wrote:
> On 10/01/09 14:05, Marco Tizzoni wrote:
>>> The clock*event* mechanism is all about setting up timers to raise events.
>>>
>>> When running paravirtualized, we use Xen-specific versions of both.
>>>
>> Ok.  My problem is granularity. In xen I can't raise events more
>> frequently than 1000/s (with timer set to 1000hz).
>> Net test softwares such as netperf or hping use setitimer to send
>> packet at a fixed rate. On a non-xen box they works fine even  with
>> 100k packets/sec (i.e. 100k alarms/sec). This can be problematic for
>> streaming software and more in general for soft real-time
>> applications.
>> [...]
>> Thanks for your test, I tried and timers work as expected.
>
> OK, good, that takes timers out of the equation.
>
> I guess the problem is the rate at which Xen will context switch between
> two domains.  Hm.
>
> Does anything change if you pin your domains to different cpus?

mmmh, no. I think without a finer timer the problem could not be solved.
For my test application, no problem, I'll use a while() loop with
counters and sched_yeld(),
but, for real application, it could be a serious issue to deal with.

m-

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel