: ANA PC router

[Frans Kaashoek]

------- Forwarded Message

To:      ibp at amsterdam.lcs.mit.edu
cc:      
Subject: ANA PC router


I spent some time chatting with jtw (Wroclawski) today.  We agreed to
meet again Friday 1p.  If you can make it, please join.  He is curious
about what we are doing; he says there is quite a bit of sceptism
about the general kind of approach we are trying.

In other news, he says that most of the PC router hacking they did a
couple of years ago made it into Linux 5.2, and that Linux would be a
fine base for us to compare to.  FreeBSD3.?  has slightly better
performance than Linux5.2, even though the Linix code is better; jtw
doesn't know why.  They are running the FreeBSD routers in production
mode, interfacing to multiple OC-3s.

He said that they completed 2 of the 3 parts of their router packet:

- - understanding what PC hardware is good (which particular
board, bus chipset, etc.); he says short packets are bottlenecked by
bus arbitration logic.  large packets are bottlenecked by the bus.
dma-ing from one interface to another network interface doesn't help
much, because most network cards don't buffer enough (or at all).

- - getting a nice fast and slow path architecture.  they didn't do any
fast IP address lookup. He says that the right code to get for IPv4
lookup is from some Swedish guys.  i don't think this matters much to
us right now.

- - they didn't finish a multiprocessor implementation: running the
router code on one dedicated node, and the rest of unix on the other
nodes.  his claim is that none of the current PC router projects are
doing this, but that this could be a big win.  his claim is if you do
this should be able to drive OC-12s.  in addition, you should be able
to get a number of other interesting properties, which he would be
happy to talk about.

	Frans


------- End of Forwarded Message