From mboxrd@z Thu Jan 1 00:00:00 1970 From: Bruce Richardson Subject: Re: Performance issue with uio_pci_generic driver Date: Fri, 18 Mar 2016 10:09:57 +0000 Message-ID: <20160318100957.GB4848@bricha3-MOBL3> References: <8379404C2611CB4BB6D4166627C18EA916BB5AB7@CMS08-02.campus.gla.ac.uk> Mime-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: quoted-printable Cc: "dev@dpdk.org" To: Simon Jouet Return-path: Received: from mga04.intel.com (mga04.intel.com [192.55.52.120]) by dpdk.org (Postfix) with ESMTP id 9D4C12C63 for ; Fri, 18 Mar 2016 11:10:16 +0100 (CET) Content-Disposition: inline In-Reply-To: <8379404C2611CB4BB6D4166627C18EA916BB5AB7@CMS08-02.campus.gla.ac.uk> List-Id: patches and discussions about DPDK List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: dev-bounces@dpdk.org Sender: "dev" On Wed, Mar 16, 2016 at 11:59:34PM +0000, Simon Jouet wrote: > Hi everyone, >=20 > First off I would like to thanks tmonjalo, Harry Van Harren and Bruce R= ichardson for the input they gave while I was trying to figure out the is= sue and pushing me to report the problem here =F0=9F=98=8A >=20 > Okay, so I was trying out some basic sanity benchmarks with DPDK before= doing anything more complicated and surprisingly I was getting lower tha= n gigabit speed for minimum packet size running l2fwd (or l3fwd for that = matter). >=20 > The setup is very simple I=E2=80=99ve got two machine with Intel x710 q= uad port NICs one is running DPDK l2fwd and the other is running MoonGen = for the performance benchmark. >=20 > After much debugging and trying to modify parameters one by one, giving= up after nothing worked and setting up ovs-dpdk I noticed from the ovs d= ocumentation that the kernel module to load were uio and igb_uio while I = was previous loading uio_pci_generic as mentioned in the DPDK getting sta= rted guide. I simply changed the kernel module and l2fwd went from 700Mbp= s to 10G line-rate. Bruce said that shouldn=E2=80=99t be the case and th= e performance should be similar regardless of the driver loaded ... >=20 > Here is the full log of the experiment, if you=E2=80=99re interested: > https://gist.github.com/simon-jouet/178e1d302afef5c6a642 >=20 > Best regards, > Simon >=20 Thanks Simon for reporting this here! It seems very strange to me that the particular module being used for uio= has such an impact on performance. Has anyone else seen anything similar? /Bruce