home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!cs.utexas.edu!swrinde!mips!sdd.hp.com!uakari.primate.wisc.edu!ames!sgi!rhyolite!vjs
- From: vjs@rhyolite.wpd.sgi.com (Vernon Schryver)
- Newsgroups: comp.dcom.lans.misc
- Subject: Re: ethernet vs. 16mb token ring
- Keywords: throughput, token ring, ethernet
- Message-ID: <oj4sv40@rhyolite.wpd.sgi.com>
- Date: 15 Aug 92 01:41:17 GMT
- References: <750@dowjone.UUCP> <odahios@rhyolite.wpd.sgi.com> <1992Aug14.184942.1790@ultra.com>
- Organization: Silicon Graphics, Inc. Mountain View, CA
- Lines: 37
-
- In article <1992Aug14.184942.1790@ultra.com>, shj@ultra.com (Steve Jay {Ultra Unix SW Mgr}) writes:
- > Hear, hear. We just got done analyzing a "problem" in which hosts
- > on an ethernet were reporting 30% "collision" rate. We were pumping
- > back-to-back packets (minimum inter-packet spacing) out of one host
- > with ttcp, and not surprisingly, every ACK packet coming back from
- > the receiving host got a "collision". That is, the receiving host always
- > tried to send its ACK at exactly the same time as the transmiting host
- > put its next packet on the wire. Each host would then back off (by
- > slightly different amounts in each host), and things would proceed
- > normally.
- >
- > It took us a while to realize that this was totally correct behavior...
- > the extraordinary collision rate being reported by the hosts was not
- > an indication of anything wrong.
- >
- > We were seeing > 1.1 Mbyte/second transfer rate between the two hosts
- > while the 30% collision rate was occuring. We get the same transfer
- > rate between other pairs of hosts without the high collision rate. We
- > decided that this likely has to do with the ethernet chips actually
- > using a value slightly different than 9.6 microseconds for the inter-
- > packet time. Just a few nanoseconds difference will prevent the
- > "collision".
-
-
- AMD LANCE 7990's have a bug wherein they are slow to start after
- deferring. They don't do anything really wrong, but they are more
- "polite" than other chips. On a "small enough" network (as measured
- in microseconds between stations), this politeness yields significantly
- better TCP/IP throughput, if you use relative large TCP windows. The
- total number of collisions is reduced, and so more time is spent moving
- data bits and less deciding whose turn it is to talk.
-
-
- Vernon Schryver, vjs@sgi.com
-
- P.S. 7990's are disappearing because other, simpler chips are easier
- and cheaper to interface to high speed buses, at least in my vicinity.
-