From mboxrd@z Thu Jan 1 00:00:00 1970 From: Phil Sutter Subject: Re: [net-next PATCH 1/3] net: make default tx_queue_len configurable Date: Thu, 13 Aug 2015 17:06:48 +0200 Message-ID: <20150813150648.GG32353@orbit.nwl.cc> References: <1438203103-27013-2-git-send-email-phil@nwl.cc> <20150729210618.GF32705@breakpoint.cc> <20150729213428.GC15554@orbit.nwl.cc> <20150729.143731.834219016906219771.davem@davemloft.net> <20150811174807.3e00b4da@redhat.com> <20150811162335.GB32353@orbit.nwl.cc> <20150812011347.GA58200@Alexeis-MacBook-Pro.local> <20150813011339.GF32353@orbit.nwl.cc> <20150813151033.048d73b8@redhat.com> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: Alexei Starovoitov , netdev@vger.kernel.org To: Jesper Dangaard Brouer Return-path: Received: from orbit.nwl.cc ([176.31.251.142]:33059 "EHLO mail.nwl.cc" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751820AbbHMPGu (ORCPT ); Thu, 13 Aug 2015 11:06:50 -0400 Content-Disposition: inline In-Reply-To: <20150813151033.048d73b8@redhat.com> Sender: netdev-owner@vger.kernel.org List-ID: On Thu, Aug 13, 2015 at 03:10:33PM +0200, Jesper Dangaard Brouer wrote: > > On Thu, 13 Aug 2015 03:13:40 +0200 Phil Sutter wrote: > > > On Tue, Aug 11, 2015 at 06:13:49PM -0700, Alexei Starovoitov wrote: > > > > In general 'changing the default' may be an acceptable thing, but then > > > it needs to strongly justified. How much performance does it bring? > > > > A quick test on my local VM with veth and netperf (netserver and veth > > peer in different netns) I see an increase of about 5% of throughput > > when using noqueue instead of the default pfifo_fast. > > Good that you can show 5% improvement with a single netperf flow. We > are saving approx 6 atomic operations avoiding the qdisc code path. > > This fixes a scalability issue with veth. Thus, the real performance > boost will happen with multiple flows and multiple CPU cores in > action. You can try with a multi core VM and use super_netperf. > > https://github.com/borkmann/stuff/blob/master/super_netperf I actually used that on my VM as well, but the difference between a single and ten streams in parallel was negligible. In order to avoid tampering the results, I tested again on a physical system with four cores, ran each benchmark ten times and built an average over the results. This showed an increase in throughput of about 35% with a single stream and about 10% with ten streams in parallel. Not sure though why the improvement is bigger in the first case if there really is a scalability problem as you say. Cheers, Phil