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

Re: [Xen-devel] A question no one can answer

  • To: "Alan Cox" <alan@xxxxxxxxxxxxxxxxxxx>
  • From: "Dan Doucette" <doucette.daniel@xxxxxxxxx>
  • Date: Fri, 15 Feb 2008 11:56:37 -0800
  • Cc: weiming <zephyr.zhao@xxxxxxxxx>, xen-devel@xxxxxxxxxxxxxxxxxxx, Robert Stober <rstober@xxxxxxxxxxxx>
  • Delivery-date: Fri, 15 Feb 2008 11:57:05 -0800
  • Domainkey-signature: a=rsa-sha1; c=nofws; d=gmail.com; s=gamma; h=message-id:date:from:to:subject:cc:in-reply-to:mime-version:content-type:references; b=x9Z2V/jBi6PiDnaKSpSa/Bt/R+6IBe1gmhMmXKFSzWrB7skrN+58QNqZ3WS7Ls5gGjxnmocNfnQ5xNGiQCL6/NstjT2NzqH/9Z/iQP/hQyN8mZCl71TRXOJ02g8O6aBuJ29MGUYvzgluwt4LDE53kC16rzVCDIbIZCUmtHLhUG8=
  • List-id: Xen developer discussion <xen-devel.lists.xensource.com>

Hello Robert and others,
Below is a link to a presentation given at a conference last summer, where the authors used cache colour allocation schemes to limit L2 cache overuse by an application. 
If your goal is to create a fair memory share per-guest (bandwidth or otherwise), you may want to consider the possibility of managing it from a level closer to the CPU.

On Fri, Feb 15, 2008 at 5:05 AM, Alan Cox <alan@xxxxxxxxxxxxxxxxxxx> wrote:
On Thu, 14 Feb 2008 22:24:18 -0500
"Robert Stober" <rstober@xxxxxxxxxxxx> wrote:

> Weiming,
> I agree that it is very hard, and that no one has done it. But
> nevertheless I suggest the following question to the Xen developers:
> Given the fact that memory bandwidth is shared amongst multiple cores on
> a single die, assume that one VM is running on each core. What is to
> stop one VM from saturating the memory bus, causing reduced performance
> of all the other VMs? This is the general multi-core problem, not
> specific to Xen. But it affects Xen greatly. What use is it to allocate
> memory to a VM if it can't use the memory because a process of another
> VM has saturated the memory bus?

Its perfectly doable on modern x86 - you use the profiling registers and
set them up so you get a threshold interrupt when too much main memory
traffic is counted (which you use to reschedule punishing the memory
user). There are research papers on it from quite a long time back but
afaik nobody ever implemented it in production although its not too hard
to do so might be an interesting project.


Xen-devel mailing list

Xen-devel mailing list



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