home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.dcom.cell-relay
- Path: sparky!uunet!mcsun!sunic!sics.se!craig
- From: craig@sics.se (Craig Partridge)
- Subject: >>>>Future of IP routers
- Message-ID: <1992Aug28.164008.4015@sics.se>
- Organization: Swedish Institute of Computer Science, Kista
- Date: Fri, 28 Aug 1992 16:40:08 GMT
- Lines: 44
-
- I'm seeing a recurrent thread of silly comments about what IP can and
- cannot do, often based on dubious benchmarks.
-
- Based on the best benchmarks I know, here's the state of the world.
-
- * Cray has documented (with careful testing) a TCP that runs
- at between 710 and 790 Mbits/s over an 800 Mbit/s HiPPI channel.
- (The difference is a matter of how you do the HiPPI signalling).
- My understanding is that this is the TCP/IP software that Cray
- sells (not some monkeyed version with the wind at its back).
-
- * The high-end routers passing IP without options can drive
- multiple FDDI interfaces full tilt (I believe the interface was
- looped such that TTRT doesn't matter). See the Harvard Router
- tests.
-
- Based on the trends (of which these results are simply the leading edge),
- there's every reason to believe that TCP can operate happily at gigabit
- speeds and that IP routers will happily forward IP datagrams at gigabit
- speeds sometime in the future. That's probably true of CLNP and TP4
- too.
-
- Kindly note that most benchmarks are extremely poor, and fail to distinguish
- between the performance of TCP/IP and the performance of interface boards
- (remember what Jacobson learned about Ethernet chips -- one ran at 10 Mbit/s
- while the other maxed out at 6 Mbit/s -- test on the wrong chip and you'd
- get the wrong value for TCP performance), local configuration issues
- (any FDDI throughput test over a real ring that doesn't mention the TTRT
- that the testers set for the network is worthless), and features of the subnet
- technology (for example, the Cray folks believe that one should release
- the HiPPI channel after each packet to allow sharing -- that takes about
- 10% off their performance). Instead, someone just runs TCP over the media
- and says "here's what it can do." If I ran an ATM performance test in which
- the receiver was intererupted for every cell and did the SAR layer
- in software, would anyone accept that as the definitive answer as to the
- maximum performance of ATM (over, say, OC-3)?
-
- Now whether you want to use IP for gigabit networks, and whether IP can
- support mixed-media traffic as well as other technologies like ATM is
- still an open question. (I happen to think that IP could work just fine
- thank-you but that's my opinion). But there's no need to make poorly
- documented comments about the performance of IP to justify ATM.
-
- Craig
-