Re: [PATCH v2] xen/credit scheduler; Use delay to control scheduling frequency

From: "Jan Beulich" <JBeulich@suse.com>
To: George Dunlap <George.Dunlap@eu.citrix.com>
Cc: Kevin Tian <kevin.tian@intel.com>,
	"xen-devel@lists.xensource.com" <xen-devel@lists.xensource.com>,
	"keir@xen.org" <keir@xen.org>, Eddie Dong <eddie.dong@intel.com>,
	Hui Lv <hui.lv@intel.com>,
	Jiangang Duan <jiangang.duan@intel.com>,
	Zhidong Yu <zhidong.yu@intel.com>
Subject: Re: [PATCH v2] xen/credit scheduler; Use delay to control scheduling frequency
Date: Mon, 19 Dec 2011 12:05:42 +0000	[thread overview]
Message-ID: <4EEF36A60200007800068D43@nat28.tlf.novell.com> (raw)
In-Reply-To: <CAFLBxZb3Yc7G4QzEXKdfrVkoz1XGv09zTxCQCC4Dnkr-6M5K5Q@mail.gmail.com>

>>> On 19.12.11 at 12:32, George Dunlap <George.Dunlap@eu.citrix.com> wrote:
> On Mon, Dec 19, 2011 at 8:24 AM, Jan Beulich <JBeulich@suse.com> wrote:
>>> +         && vcpu_runnable(current)
>>> +         && !is_idle_vcpu(current)
>>> +         && runtime < MICROSECS(sched_ratelimit_us) )
>>> +    {
>>> +        snext = scurr;
>>> +        snext->start_time += now;
>>> +        perfc_incr(delay_ms);
>>> +        tslice = MICROSECS(sched_ratelimit_us);
>>
>> So if there happens to be a VM with
>>
>> MILLISECS(prv->tslice_ms) < MICROSECS(sched_ratelimit_us)
>>
>> it'd get *more* time than allowed/intended through this mechanism.
> 
> Yeah, if you set your default timeslice to 1ms, and then set your
> minimum scheduling rate to 5ms, you're going to get weird results. :-)
> 
> The way it stands now, the ratelimit value will override the timeslice
> value.  It had to be one way or the other; do you think the timeslice
> value should be the priority?

The minimum of both should be used, I would think.

Jan