Date: Wed, 04 Jul 2012 12:46:09 +0400 From: "Alexander V. Chernikov" <melifaro@FreeBSD.org> To: Doug Barton <dougb@FreeBSD.org> Cc: net@freebsd.org Subject: Re: FreeBSD 10G forwarding performance @Intel Message-ID: <4FF402D1.4000505@FreeBSD.org> In-Reply-To: <4FF3FB14.8020006@FreeBSD.org> References: <4FF319A2.6070905@FreeBSD.org> <4FF361CA.4000506@FreeBSD.org> <20120703214419.GC92445@onelab2.iet.unipi.it> <4FF36438.2030902@FreeBSD.org> <4FF3E2C4.7050701@FreeBSD.org> <4FF3FB14.8020006@FreeBSD.org>
next in thread | previous in thread | raw e-mail | index | archive | help
On 04.07.2012 12:13, Doug Barton wrote: > On 07/03/2012 23:29, Alexander V. Chernikov wrote: >> On 04.07.2012 01:29, Doug Barton wrote: >>>>> Just curious ... what's the MTU on your FreeBSD box, and the Linux box? >> >> In this particular setup - 1500. You're probably meaning type of mbufs >> which are allocated by ixgbe driver? > > 1500 for both? Well, AFAIR it was 1500. We've done a variety of tests half a year ago with similar server and Intel and Mellanox equipment. Test results vary from 4 to 6mpps in different setups (and mellanox seems to behave better on Linux). If you're particularly interested in exact Linux performance on exactly the same box I can try to do this possibly next week. My point actually is the following: It is possible to do linerate 10G (14.8mpps) forwarding with current market-available hardware. Linux is going that way and it is much more close than we do. Even dragonfly performs _much_ better than we do in routing. http://shader.kaist.edu/packetshader/ (and links there) are good example of what is going on. > > And no, I'm not thinking of the mbufs directly, although that may be a > side effect. I've seen cases on FreeBSD with em where setting the MTU to > 9000 had unexpected (albeit pleasant) side effects on throughput vs. Yes. Stock drivers has this problem, especially with IPv6 addresses. We actually use our versions of em/igb/ixgbe drivers in production which are free from several problems in stock driver. (Tests, however, were done using stock driver) > system load. Since it was working better I didn't take the time to find > out why. However since you're obviously interested in finding out the > nitty-gritty details (and thank you for that) you might want to give it > a look, and a few test runs. > > hth, > > Doug >
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?4FF402D1.4000505>