home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.dcom.cell-relay
- Path: sparky!uunet!munnari.oz.au!mips!mips!sgi!rhyolite!vjs
- From: vjs@rhyolite.wpd.sgi.com (Vernon Schryver)
- Subject: Re: Future of IP routers
- Message-ID: <p6dd90s@rhyolite.wpd.sgi.com>
- Organization: Silicon Graphics, Inc. Mountain View, CA
- References: <1992Aug26.092945.4663@ccsun.strath.ac.uk> <3445@ra.nrl.navy.mil> <1992Aug28.151702.12575@advtech.uswest.com>
- Date: Sat, 29 Aug 1992 16:24:49 GMT
- Lines: 57
-
- In article <1992Aug28.151702.12575@advtech.uswest.com>, jsims@uswest.com (John Sims) writes:
-
- > Could you be more specific about the environment in which these results
- > were acheived and exactly what IP over ATM results you expect?
- > I have one of the early implementations of ATM for Suns and I see
- >throughput (Sparc2 to Sparc2) of 19.5Mbps TCP/IP over a 100 Mbps ATM interface.
- > So the results reported by M. Burak seem in-line with my observations.
- > Not that someone couldn't produce a better implementation, so I'm asking if
- > anyone out there knows of one?
-
- At least some of the current crop of ATM cards for workstations
- move all data with "programmed I/O". In other words, the card
- is not a DMA master. That they do better than ethernet amazes me.
- Their speed has nothing to do with whatever ATM will eventually provide.
-
-
- > It seems obvious to me that (all other factors being equal) TCP/IP throughput
- > on an FDDI would be better than throughput on a 100Mbps ATM. And of course
- > I'm talking about just having 2 stations on the FDDI, etc... The overhead
- > that ATM adds to be able to support voice and video would make it less
- > efficient at carrying TCP/IP than an FDDI. Anyone disagree?
-
- In the simplest, unintended sense, an ATM switch can be many times
- faster than an FDDI ring. Each pair of ATM ports can be running at
- 155Mb/s (or whatever), but you get only 100Mb/s (or 200Mb/s if you
- believe in dual-MAC) to divide among all stations on an FDDI ring.
-
- I disagree with the intended point in the weak sense of being unwilling
- to entertain an opinion yet. Few of the few people who have ever moved
- TCP data faster than 50Mbit/sec over any medium say much of anything to
- anyone. Many people who speculate about commercially available network
- speeds above 50Mbit/sec would be surprised by what is really involved,
- by the real nature of the bottlenecks.
-
- ATM, as I understand it, can be more full-duplex than FDDI. So you
- might argue that 100 Mb ATM would be faster. Then again, those 53 byte
- cells waste a lot of bandwidth. On the other hand, you don't hear so
- much about 100Mb/s ATM these days compared to 155 or 165 or other
- rates, and 155Mb ATM sounds faster than 100Mb FDDI. However,
- assempling cells into IP packets is not going to be cheap and could
- introduce significant latency, and latency is Not Your Friend when
- you're trying to push a single virtual circut. On the other hand, FDDI
- latencies can be amazing.
-
- I have spent some time trying to make TCP run fast over FDDI. I have
- learned just enough to know that "it just aint that simple," at least
- for me.
-
- For example, if your FDDI station uses medium sized TCP windows
- [32K,64K) and can easily transmit at > 100 Mb/s, then you might find it
- difficult to reach 98Mb/s because of the half-duplex, release the token
- only nothing more is available to transmit nature of FDDI. (Consider
- when the ACKs from the receiver are seen by the transmitter, and the
- potential idle time on the ring.)
-
-
- Vernon Schryver, vjs@sgi.com
-