oh, no, forwarding rates are not very good. I have a bunch of numbers on forwarding rates on the wndr3800 stuff, at best it's half that for routing rather than bridging. I can give you current pps or bytes/sec, but not right now. (felix just did a ton of tuning in this area). Bridging outperforms routing by like 6x1.
Add in firewall rules or nat, or a software rate limiter, and you are well below 100Mbit for forwarding rates over the ethernet on that particular box. It's not the SOC per se' that's the issue, it's the tiny caches (vs, like, um, ivy bridge), and the (generally 16 bit single channel) pretty slow memory interfaces common in these platforms, as well as MIP's general decrepitude. On Wed, Aug 14, 2013 at 8:13 AM, Jim Gettys <[email protected]> wrote: > > > > On Mon, Aug 12, 2013 at 5:40 AM, Mikael Abrahamsson <[email protected]>wrote: > >> On Mon, 12 Aug 2013, Teco Boot wrote: >> >> Joel Jaeggli mentioned the forwarding performance. Today's homenets are >>> mainly single subnet with link-speed forwarding between (gig) ethernet >>> ports, in hardware. L3 forwarding in software with single uplink or WiFi >>> port at near to gig speed is doable. Forwarding in software on all ports >>> require a new generation of low power and cheap CPU, I think. So probably >>> use hardware forwarding as much as possible? >>> >> >> Hardware assisted forwarding might be problematic due to us deciding on >> new functionality (source based routing for instance). I've read that in >> some routers the forwarding is done by microcode implemented by the >> hardware manufacturer, hindering the integrator (who buys the SoC in bulk) >> from doing what might be needed. >> >> So yes, forwarding performance is a concern, at least when we're talking >> above 100 megabit/s. >> > > Actually, I think (Dave Taht can tell us with data), that forwarding > performance is fine even on today's routers up to 300-400Mbps on the > WNDR3800 running CeroWrt (which routes, rather than bridging between > interfaces).. The 600Mhz MIPS can't saturate a gigabit ethernet, however > (at least not without using lots of hardware offload on the ethernet > controller, which hurts latency and creates > other problems (packet bursts). > > But as SOC's are increasing in speed while the price goes down... > >> >> I also think it would be beneficial if we could figure out as soon as >> possible what the requirements are on the forwarding plane, writing this >> down, so that hardware designers can avoid putting functionality into >> hardware that won't do what we need going fo >> r >> ward. > > > Hardware "assists", e.g. TSO, and GSO, are a problem, particularly in the > home. You don't want line rate bursts going onto your WiFi (though > fq_codel helps that quite a bit, it's still evil). > >> >> >> -- >> Mikael Abrahamsson email: [email protected] >> ______________________________**_________________ >> homenet mailing list >> [email protected] >> https://www.ietf.org/mailman/**listinfo/homenet<https://www.ietf.org/mailman/listinfo/homenet> >> > > > _______________________________________________ > homenet mailing list > [email protected] > https://www.ietf.org/mailman/listinfo/homenet > > -- Dave Täht Fixing bufferbloat with cerowrt: http://www.teklibre.com/cerowrt/subscribe.html
_______________________________________________ homenet mailing list [email protected] https://www.ietf.org/mailman/listinfo/homenet
