Date: Wed, 14 Jun 2017 11:44:35 -0500 From: Jim Thompson <jim@netgate.com> To: John Jasen <jjasen@gmail.com> Cc: FreeBSD Net <freebsd-net@freebsd.org> Subject: Re: state of packet forwarding in FreeBSD? Message-ID: <64C72460-54F3-481B-9A2C-044822F6F52B@netgate.com> In-Reply-To: <CAACLuR17yRETErqsxbdhBPJrjQur0oMVOqvL5ZCkmjLCKkHLNA@mail.gmail.com> References: <CAACLuR17yRETErqsxbdhBPJrjQur0oMVOqvL5ZCkmjLCKkHLNA@mail.gmail.com>
next in thread | previous in thread | raw e-mail | index | archive | help
> On Jun 14, 2017, at 9:48 AM, John Jasen <jjasen@gmail.com> wrote: >=20 > Our goal was to test whether or not FreeBSD currently is viable, as = the > operating system platform for high speed routers and firewalls, in the > 40 to 100 GbE range. We recently showed IPsec running at 36.32Gbps (8 streams, 32.68Gbps = single stream). =20 At 36.32Gbps, we were limited by the 40Gbps cards we used. (Framing overheads take up about 10% of the bandwidth.) = https://conferences.oreilly.com/oscon/oscon-tx/public/schedule/detail/5672= 7 = <https://conferences.oreilly.com/oscon/oscon-tx/public/schedule/detail/567= 27> We can send 64 byte tiny grams through the tunnel at 10.45 Mpps with = AES-128-CBC + HMAC-SHA1. AES-128-GCM performance is 32.98Gbps (4 = streams, 32.72Gbps single-stream). Hardware used was essentially an =E2=80=9Cultimate white box router=E2=80=9D= : https://www.netgate.com/blog/building-a-behemoth-router.html = <https://www.netgate.com/blog/building-a-behemoth-router.html>; with = Intel xl710 NICs and 8955 CPIC QAT cards. The same hardware will l3 forward 42.6Mpps (64 byte packets). It can = forward 14.05Mpps on a single core. No tuning was done in the above, = just bringing up VPP configuring the interfaces and SPDs, and running = iperf3 or (DPDK=E2=80=99s) pkggen on a pair of outside =E2=80=98hosts=E2=80= =99. It=E2=80=99s likely that we can get the 42.6Mpps figure higher. In other tests on smaller (Atom 8 core) hardware we=E2=80=99ve achieved = 12Mpps l3 forwarding with a full BGP routing table. Likely that we can = achieve even higher PPS results with a bit more tuning work. Using Olivier Couchard-Labb=C3=A9=E2=80=99s =E2=80=9Cestimated IMIX=E2=80=9D= (PPS * ( 7*(40+14) + 4*(576+14) + (1500+14) )/12*8), one only needs = 27Mpps to fill a 100Gbps interface. We have a couple larger machines with unreleased Xeons in them, 4 = 100gbps NICs, and some =E2=80=9Cnext generation=E2=80=9D QAT cards that = Intel says are good for 100Gbps encryption offload. We plan to re-run = the tests at 100Gbps sometime this summer. These results are all on Linux, using VPP over DPDK, but nothing really = restricts that work from moving back to FreeBSD. VPP also supports = netmap, but we=E2=80=99ve not attempted any performance work using the = netmap interfaces as yet. = https://fd.io/news/announcement/2017/06/fast-data-fdio-project-issues-four= th-release-furthers-position-universal = <https://fd.io/news/announcement/2017/06/fast-data-fdio-project-issues-fou= rth-release-furthers-position-universal> gnn@ was working on such a port for us, but other things took over his = time. https://github.com/gvnn3/vpp <https://github.com/gvnn3/vpp>. = I=E2=80=99m sure we=E2=80=99ll get back to it. (This is all the basel for our =E2=80=9Cnext generation pfSense=E2=80=9D, = btw.) > In our investigations, we tested 10.3, 11.0/-STABLE, -CURRENT, and a = USB > stick from BSDRP using the FreeBSD routing improvements project > enhancements (https://wiki.freebsd.org/ProjectsRoutingProposal). >=20 > We've tried stock and netmap-fwd, have played around a little with > netmap itself and dpdk, with the results summarized below. The current > testing platform is a Dell PowerEdge R530 with a Chelsio T580-LP-CR = dual > port 40GbE card. >=20 > Suggestions, examples for using netmap, etc, all warmly welcomed. >=20 > Further questions cheerfully answered to the best of our abilities. >=20 > a) On the positive side, it appears that 11.0 is much faster than = 10.0, > which we tested several years ago. With appropriate cpuset tuning, 5.5 > mpps is achievable using modern hardware. Using slightly older = hardware, > (such as a Dell R720 with v3 xeons), around 5.2-5.3 mpps can be = obtained. >=20 > b) On the negative side, between the various releases, netmap appeared > to be unstable with the Chelsio cards -- sometimes supported, = sometimes > broken. Also, we're still trying to figure out netmap utilities, such = as > vale-ctl and bridge, so any advice would be appreciated. >=20 > b.1) netmap-fwd is admittedly single-threaded and does not support = IPv6. There is a version of netmap-fwd (not on GitHub) that supports IPv6, and = has some early work on threading. Unfortunately netmap bugs stopped the threading work. The developer (loos@) recently updated netmap in -CURRENT based on a = patch from Vincenzo Maffione. BTW, we=E2=80=99ve seen over 5mpps using netmap-fwd using a (single core = of a) E3-1275. See around 17:07 = https://www.youtube.com/watch?v=3DcAVgvzivzII > These clearly showed in our tests, as we were unable to achieve over = 2.5 > mpps, saturating a single CPU and letting the others fall asleep. > However, bumping a single CPU queue from around 0.6 mpps to 2.5 mpps = is > nothing to ignore, so it could be useful in some cases. >=20 > c) The routing improvement project USB stick performed incredibly, > achieving 8.5 mpps out of the box. However, it appears > (https://wiki.freebsd.org/ProjectsRoutingProposal/ConversionStatus), > that many of the changes are still pending review, and that things = have > not moved much in the last 18 months > (https://svnweb.freebsd.org/base/projects/routing/) >=20 > d) We've not figured out dpdk (dpdk.org) yet. Our first foray into = the > test examples, and we're stuck trying to get the interfaces online. DPDK on FreeBSD is a bit of a mess. Jim =20=
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?64C72460-54F3-481B-9A2C-044822F6F52B>