home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.dcom.cell-relay
- Path: sparky!uunet!mcsun!dxcern!brian
- From: brian@dxcern.cern.ch (Brian Carpenter CERN-CN)
- Subject: Re: Congestion Avoidance
- Message-ID: <1992Jul27.101327.12625@dxcern.cern.ch>
- Organization: CERN European Lab for Particle Physics
- References: <1992Jul18.230550.3046@sics.se> <1992Jul19.130005.1@research.ptt.nl> <1992Jul25.092823.12189@e2big.mko.dec.com> <k3jm90-.nagle@netcom.com>
- Date: Mon, 27 Jul 1992 10:13:27 GMT
- Lines: 33
-
- In <k3jm90-.nagle@netcom.com> nagle@netcom.com (John Nagle) writes:
-
- > Uh oh. As the one-time congestion maven of the Internet
- >(look at the RFCs that bear my name) I somehow thought the cell relay
- >people had developed a new solution to the problem that would make
- >huge connectionless nets work. If they haven't, there's going to be
- >trouble a few years downstream.
-
- There is nothing new under the sun and M/D/n queuing systems still
- behave the same as when Kleinrock wrote the book. The most convincing
- story I've heard from an ATM switch fabric designer is "we've ensured
- that the fabric can saturate all its outputs simultaneoulsy, and we've
- put in enough buffers to guarantee correct re-ordering of cells at every
- output." What this means is that if the total traffic at the _inputs_
- exceeds the total capacity of the outputs for longer than some delta-T
- you are, surprise surprise, in trouble. Of course the designers have done
- statistical simulations, but they remain very scared of connectionless
- data traffic.
-
- > The existing (NSFNET) approach of generously sizing the backbone using
- >government subsidies, may not translate well to the commercial world.
-
- Well, as long as your switching fabric is designed to be scalable you
- can improve switch performance up to a point just by adding silicon -
- and this seems to be more crucial than the bandwidth of trunk lines,
- in controlling cell loss rates. (ATM is actually connection-oriented
- with guaranteed bandwidth per channel.) But in the end, the queuing theory
- will get you.
-
-
- Regards,
- Brian Carpenter CERN, brian@dxcern.cern.ch
- voice +41 22 767 4967, fax +41 22 767 7155
-