Tony Li writes:
> BTW, NetBSD with the recent flow cache mods can handle at least
> 150,000 packets per second. We haven't seen what the actual upper
> limit is, but that number doesn't seem to be eating a lot of CPU.
Interesting. Have you done any scalability testing? Per-flow state has
been shown to scale poorly in Internet backbones.
I agree with you that it doesn't scale well to huge routers. I am
certain that it would work very badly in MCI's backbone. It probably
will work just fine in a small provider, though, and those are most of
the guys who want to roll their own out of PCs anyway.
If you want to pump half a gigabit through on a giant peering router,
caches will lose, and the only thing that will do you right is
real router hardware.
Perry
perry@piermont.com (Perry E. Metzger) writes:
> Interesting. Have you done any scalability testing? Per-flow state has
> been shown to scale poorly in Internet backbones.
I agree with you that it doesn't scale well to huge routers. I am
certain that it would work very badly in MCI's backbone. It probably
will work just fine in a small provider, though, and those are most of
the guys who want to roll their own out of PCs anyway.
Not a foregone conclusion. Recall that it's not the absolute volume of the
traffic that's the difficulty but the diversity and 'hit ratio' of the
particular traffic patterns. If the flow cache is small and stable, then
you could push terabits. The problem that has been seen is that anything
connected to the Internet with a significant community behind it has
critical mass to churn the flow cache.
If you want to pump half a gigabit through on a giant peering router,
caches will lose, and the only thing that will do you right is
real router hardware.
I wouldn't know anything about that.
Tony