From mboxrd@z Thu Jan 1 00:00:00 1970 From: Jambunathan K Subject: Re: PCI Passthrough & iperf Throughput Skew Date: Tue, 27 Mar 2007 20:10:17 +0530 Message-ID: <46092CD1.3090903@netxen.com> References: <46092BBA.6000105@netxen.com> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit Return-path: In-Reply-To: <46092BBA.6000105@netxen.com> List-Unsubscribe: , List-Post: List-Help: List-Subscribe: , Sender: xen-devel-bounces@lists.xensource.com Errors-To: xen-devel-bounces@lists.xensource.com To: xen-devel@lists.xensource.com Cc: Sanjeev Jorapur List-Id: xen-devel@lists.xenproject.org Additional Data Point: If I just exercise iperf in Dom0 or DomU *separately*, I get identical throughput . The 10:1 skew happens only when the iperf is running concurrently on the Dom0 and DomU. Regards, Jambunathan K. Jambunathan K wrote: > I am seeing a strange behaviour with Xen-3.0.4 which seems to suggest > that Dom0 has either higher Scheduling affinity or Interrupt affinity. > > I am in the initial stages of investigation and it is possible that my > hunch is totally wrong. I would like to hear comments or questions > from the experts. > > I have a multifunction NIC card, with one function being used by Dom0 > and another one exported to DomU using PCI passthrough > > Dom0 and DomU have almost the same memory allocated to each > (~1.3G). The ports themselves are operating at the same speed and > share the same IRQ. > > iperf (as a client) is started on Dom0 in the same way as it is > started on DomU. > > iperf results are skewed in favour of Dom0 in the ratio 10:1. > > Given all other things are similar wouldn't it be normal to expect > throughputs in the ratio of 1:1. > > (ps:- The 2 ports on the Xen machine is connected > "point-to-point-wise" to 2 ports on another machine running Linux > natively) > > > Regards, > Jambunathan K. > > > > > > > >