Re: Ideas Re: [PATCH v14 1/2] vmx: VT-d posted-interrupt core logic handling

From: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
To: Dario Faggioli <dario.faggioli@citrix.com>
Cc: Kevin Tian <kevin.tian@intel.com>, Keir Fraser <keir@xen.org>,
	George Dunlap <George.Dunlap@eu.citrix.com>,
	Andrew Cooper <andrew.cooper3@citrix.com>,
	"xen-devel@lists.xen.org" <xen-devel@lists.xen.org>,
	Jan Beulich <jbeulich@suse.com>, Feng Wu <feng.wu@intel.com>
Subject: Re: Ideas Re: [PATCH v14 1/2] vmx: VT-d posted-interrupt core logic handling
Date: Mon, 7 Mar 2016 15:23:52 -0500	[thread overview]
Message-ID: <20160307202351.GC18590@char.us.oracle.com> (raw)
In-Reply-To: <1457367599.3102.41.camel@citrix.com>

On Mon, Mar 07, 2016 at 05:19:59PM +0100, Dario Faggioli wrote:
> On Mon, 2016-03-07 at 10:53 -0500, Konrad Rzeszutek Wilk wrote:
> > On Mon, Mar 07, 2016 at 11:21:33AM +0000, George Dunlap wrote:
> > > 
> > > > <handwaving>
> > > > Would it be perhaps possible to have an anti-affinity flag to
> > > > deter the
> > > > scheduler from this? That is whichever struct vcpu has 'anti-
> > > > affinity' flag
> > > > set - the scheduler will try as much as it can _to not_ schedule
> > > > the 'struct vcpu'
> > > > if the previous 'struct vcpu' had this flag as well on this pCPU?
> >
> That can also be seen as step in the direction of (supporting) gang
> scheduling, which we've said already it would be something interesting
> to look at, although difficult to implement and even more difficult to
> figure out whether it is actually a good thing for most workloads.
> 
> In any case, I see from where this comes, and am up for thinking about
> it, although my fear is that it would complicate the code by quite a
> bit, so I agree with George that profiling work is necessary to try to
> assess whether it could be really useful (as well as, once
> implemented/drafted, whether it is really good and does not cause perf
> regressions).
> 
> > > On the whole it seems unlikely that having two vcpus on a single
> > > pcpu
> > > is a "stable" situation -- it's likely to be pretty transient, and
> > > thus not have a major impact on performance.
> > Except that we are concerned with it - in fact we are disabling this
> > feature because it may happen. 
> >
> I'm sorry, I'm not getting, what feature are you disabling?

It is already disabled in the code:

 62 /*                                                                              
 63  * In the current implementation of VT-d posted interrupts, in some extreme     
 64  * cases, the per cpu list which saves the blocked vCPU will be very long,      
 65  * and this will affect the interrupt latency, so let this feature off by       
 66  * default until we find a good solution to resolve it.                         
 67  */                                                                             
 68 bool_t __read_mostly iommu_intpost;                     

> 
> > > But I think some profiling is in order before anyone does serious
> > > work on this.
> > I appreciate your response being 'profiling' instead of 'Are you
> > NUTS!?' :-)
> > 
> That's only because everyone knows you're nuts, there's no need to
> state it all the times! :-P :-P

<laughs>

Glad that you have the _right_ expectations of me :) 

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
http://lists.xen.org/xen-devel