[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[Xen-devel] Re: [PATCH]x86: add a range for hpet broadcast


  • To: "Wei, Gang" <gang.wei@xxxxxxxxx>, Xen-devel <xen-devel@xxxxxxxxxxxxxxxxxxx>
  • From: Keir Fraser <keir.fraser@xxxxxxxxxxxxx>
  • Date: Wed, 09 Dec 2009 13:09:39 +0000
  • Cc:
  • Delivery-date: Wed, 09 Dec 2009 05:10:07 -0800
  • List-id: Xen developer discussion <xen-devel.lists.xensource.com>
  • Thread-index: Acp4sqXoSopeIMedTaiA9e0L2HFpxQADZH79AAG+pBAAAmqknQ==
  • Thread-topic: [PATCH]x86: add a range for hpet broadcast

Thanks. Also, beyond a certain point, you have to consider whether just
setting timer_slop=2ms (or whatever) has similar effect on performance and
power saving, with no code modifications. :-) It's not like slop=1ms is some
super-special value or anything, arrived at with scientific exactitude.

But the approach I suggect at least seems internally consistent and I will
considewr it if it gives you similar power wins to what your original patch
gave you.

 -- Keir

On 09/12/2009 12:34, "Wei, Gang" <gang.wei@xxxxxxxxx> wrote:

> On a Xeon 5500 server w/ 2S/8C/16T, w/ HT off, starting 8 rhel5 UP hvm guests,
> idling, I saw 22W (10%) power saving after apply this patch and specify
> timer_slop=1ms on xen grub line. It saves 10W more than the case only specify
> timer_slop=1ms but w\o this patch.
> 
> Yes, your suggestion do remove the double dipping. I will give it a try.
> 
> Jimmy
> 
> Keir Fraser wrote:
>> How much of a win is this? The per-cpu timer_deadline is already
>> calculated based on slop, so doing it again in hpet.c is kind of like
>> double dipping. I think we can validly do some improvements in hpet.c
>> however: e.g., by exposing a per-cpu timer deadline range
>> (deadline_start,deadline_end) and setting up HPET to fire on the
>> earliest deadline_end. Then broadcast to all CPUs with deadline_start
>> < NOW() when the HPET fires.
>> 
>> The deadline range would be computed as follows in timer.c, in the
>> !ts->overflow case in the softirq handler:
>> timer_deadline_start = start (i.e., same as timer_deadline today)
>> timer_deadline_end = end
>> In the (incredibly uncommon) ts->overflow case we can just set
>> timer_deadline_start=timer_deadline_end=timer_deadline.
>> 
>> Perhaps give that a go: I would find that more acceptable at least,
>> but all of this really depends on what benchmark improvements you are
>> seeing. 
>> 
>>  -- Keir
>> 
>> On 09/12/2009 09:33, "Wei, Gang" <gang.wei@xxxxxxxxx> wrote:
>> 
>>> x86: add a range for hpet broadcast
>>> 
>>> Apply a range timer like range to hpet broadcast, so that timer
>>> expires within timer_slop ns across idle CPUs are capable to be
>>> aligned to reduce hpet intrs, and save idle power.
>>> 
>>> Signed-off-by: Wei Gang <gang.wei@xxxxxxxxx>



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


 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.