From mboxrd@z Thu Jan 1 00:00:00 1970 From: David Xu Subject: Re: Re: performance of credit2 on hybrid workload Date: Tue, 7 Jun 2011 15:28:52 -0400 Message-ID: References: <1306340309.21026.8524.camel@elijah> <1306401493.21026.8526.camel@elijah> Mime-Version: 1.0 Content-Type: multipart/mixed; boundary="===============0067724375==" Return-path: In-Reply-To: List-Unsubscribe: , List-Post: List-Help: List-Subscribe: , Sender: xen-devel-bounces@lists.xensource.com Errors-To: xen-devel-bounces@lists.xensource.com To: George Dunlap , xen-devel@lists.xensource.com List-Id: xen-devel@lists.xenproject.org --===============0067724375== Content-Type: multipart/alternative; boundary=000e0cd644c0eecf4704a5243a0b --000e0cd644c0eecf4704a5243a0b Content-Type: text/plain; charset=ISO-8859-1 Hi George, Could you share some ideas about how to addressed the "mixed workload" problem, where a single VM does both cpu-intensive and latency-sensitive workloads, even though you haven't implemented it yet? I am also working on it, maybe I can try some methods and give you feedback. Thanks. Regards, Cong 2011/6/1 George Dunlap > You cannot do that with the current code; to add such a parameter > would require major work to the scheduler. > > -George > > On Wed, Jun 1, 2011 at 1:55 AM, David Xu wrote: > > Hi, > > I want to reduce the latency of a specific VM. How should I do based on > > credit scheduler? For example, I will add another parameter latency > besides > > weight and cap, and schedule the vcpu whose VM holds the least latency > > firstly each time. Thanks. > > Regards, > > Cong > > > > 2011/5/26 George Dunlap > >> > >> Please reply to the list. :-) > >> > >> Also, this is a question about credit1, so it should arguably be a > >> different thread. > >> > >> -George > >> > >> On Wed, 2011-05-25 at 19:34 +0100, David Xu wrote: > >> > Thanks. The boost mechanism in credit can significantly reduce the > >> > scheduling latency for pure I/O workload. Since the minimum interval > >> > of credit scheduling is 10ms, the magnitude of latency for the target > >> > VM should be 10ms (except the credit is not used up and vcpu remain > >> > the head of runqueue ) as well. Why the real latency in my test (Ping > >> > the target VM) is much shorter than 10ms? Does the vcpu of target VM > >> > remain the head of runqueue if it was boosted? > >> > > >> > > >> > David > >> > > >> > 2011/5/25 George Dunlap > >> > > >> > On Mon, 2011-05-23 at 09:15 +0100, David Xu wrote: > >> > > Hi, > >> > > > >> > > > >> > > Xen4.1 datasheet tells that credit2 scheduler is designed > >> > for latency > >> > > sensitive workloads. Does it have some improvement on the > >> > hybrid > >> > > workload including both the cpu-bound and latency-sensitive > >> > i/o work? > >> > > For example, if a VM runs a cpu-bound task burning the cpu > >> > and a > >> > > i/o-bound (latency-sensitive) task simultaneously, will the > >> > latency be > >> > > guaranteed? And how? > >> > > >> > > >> > At the moment, the "mixed workload" problem, where a single VM > >> > does both > >> > cpu-intensive and latency-sensitive* workloads, has not been > >> > addressed > >> > yet. I have some ideas, but I haven't implemented them yet. > >> > > >> > * i/o-bound is not the same as latency sensitive. They > >> > obviously go > >> > together frequently, but I would make a distinction between > >> > them. For > >> > example, an scp (copy over ssh) can easily become cpu-bound if > >> > there is > >> > competition for the cpu -- but it is nonetheless latency > >> > sensitive. (I > >> > guess to put it another way, a workload which is > >> > latency-sensitive may > >> > become i/o-bound if its scheduling latency is too high.) > >> > > >> > -George > >> > > >> > > >> > > >> > >> > > > > > > _______________________________________________ > > Xen-devel mailing list > > Xen-devel@lists.xensource.com > > http://lists.xensource.com/xen-devel > > > > > --000e0cd644c0eecf4704a5243a0b Content-Type: text/html; charset=ISO-8859-1 Content-Transfer-Encoding: quoted-printable Hi George,

Could you share some ideas about how to addre= ssed the=A0=A0"mixed work= load" problem,=A0= =A0where a single VM does both
cpu-intensive and latency-sensitive workloads, even though you haven't = implemented it yet? =A0I am also working on it, maybe I can try some method= s and give you feedback. Thanks.

Regards,
Cong



2011/6= /1 George Dunlap <George.Dunlap@eu.citrix.com>
You cannot do that with the current code; to add such a parameter
would require major work to the scheduler.

=A0-George

On Wed, Jun 1, 2011 at 1:55 AM, David Xu <davidxu06@gmail.com> wrote:
> Hi,
> I want to reduce the latency of a specific VM. How should I do based o= n
> credit scheduler? For example, I will add another parameter latency be= sides
> weight and cap, and schedule the vcpu whose VM holds the least latency=
> firstly each time.=A0Thanks.
> Regards,
> Cong
>
> 2011/5/26 George Dunlap <george.dunlap@citrix.com>
>>
>> Please reply to the list. :-)
>>
>> Also, this is a question about credit1, so it should arguably be a=
>> different thread.
>>
>> =A0-George
>>
>> On Wed, 2011-05-25 at 19:34 +0100, David Xu wrote:
>> > Thanks. The boost mechanism in credit can significantly reduc= e the
>> > scheduling latency for pure I/O workload. Since the minimum i= nterval
>> > of credit scheduling is 10ms, the magnitude of latency for th= e target
>> > VM should be 10ms (except the credit is not used up and vcpu = remain
>> > the head of runqueue ) as well. Why the real latency in my te= st (Ping
>> > the target VM) is much shorter than 10ms? Does the vcpu of ta= rget VM
>> > remain the head of runqueue if it was boosted?
>> >
>> >
>> > David
>> >
>> > 2011/5/25 George Dunlap <george.dunlap@citrix.com>
>> >
>> > =A0 =A0 =A0 =A0 On Mon, 2011-05-23 at 09:15 +0100, David Xu w= rote:
>> > =A0 =A0 =A0 =A0 > Hi,
>> > =A0 =A0 =A0 =A0 >
>> > =A0 =A0 =A0 =A0 >
>> > =A0 =A0 =A0 =A0 > Xen4.1 datasheet tells that credit2 sche= duler is designed
>> > =A0 =A0 =A0 =A0 for latency
>> > =A0 =A0 =A0 =A0 > sensitive workloads. Does it have some i= mprovement on the
>> > =A0 =A0 =A0 =A0 hybrid
>> > =A0 =A0 =A0 =A0 > workload including both the cpu-bound an= d latency-sensitive
>> > =A0 =A0 =A0 =A0 i/o work?
>> > =A0 =A0 =A0 =A0 > For example, if a VM runs a cpu-bound ta= sk burning the cpu
>> > =A0 =A0 =A0 =A0 and a
>> > =A0 =A0 =A0 =A0 > i/o-bound (latency-sensitive) task simul= taneously, will the
>> > =A0 =A0 =A0 =A0 latency be
>> > =A0 =A0 =A0 =A0 > guaranteed? And how?
>> >
>> >
>> > =A0 =A0 =A0 =A0 At the moment, the "mixed workload"= problem, where a single VM
>> > =A0 =A0 =A0 =A0 does both
>> > =A0 =A0 =A0 =A0 cpu-intensive and latency-sensitive* workload= s, has not been
>> > =A0 =A0 =A0 =A0 addressed
>> > =A0 =A0 =A0 =A0 yet. =A0I have some ideas, but I haven't = implemented them yet.
>> >
>> > =A0 =A0 =A0 =A0 * i/o-bound is not the same as latency sensit= ive. =A0They
>> > =A0 =A0 =A0 =A0 obviously go
>> > =A0 =A0 =A0 =A0 together frequently, but I would make a disti= nction between
>> > =A0 =A0 =A0 =A0 them. =A0For
>> > =A0 =A0 =A0 =A0 example, an scp (copy over ssh) can easily be= come cpu-bound if
>> > =A0 =A0 =A0 =A0 there is
>> > =A0 =A0 =A0 =A0 competition for the cpu -- but it is nonethel= ess latency
>> > =A0 =A0 =A0 =A0 sensitive. =A0(I
>> > =A0 =A0 =A0 =A0 guess to put it another way, a workload which= is
>> > =A0 =A0 =A0 =A0 latency-sensitive may
>> > =A0 =A0 =A0 =A0 become i/o-bound if its scheduling latency is= too high.)
>> >
>> > =A0 =A0 =A0 =A0 =A0-George
>> >
>> >
>> >
>>
>>
>
>
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xenso= urce.com
> htt= p://lists.xensource.com/xen-devel
>
>

--000e0cd644c0eecf4704a5243a0b-- --===============0067724375== Content-Type: text/plain; charset="us-ascii" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit Content-Disposition: inline _______________________________________________ Xen-devel mailing list Xen-devel@lists.xensource.com http://lists.xensource.com/xen-devel --===============0067724375==--