home *** CD-ROM | disk | FTP | other *** search
- Xref: sparky comp.dcom.cell-relay:425 comp.dcom.lans.fddi:575
- Newsgroups: comp.dcom.cell-relay,comp.dcom.lans.fddi
- Path: sparky!uunet!usc!elroy.jpl.nasa.gov!ames!sgi!rhyolite!vjs
- From: vjs@rhyolite.wpd.sgi.com (Vernon Schryver)
- Subject: Re: >>>>Future of IP routers
- Message-ID: <p87dj5o@rhyolite.wpd.sgi.com>
- Organization: Silicon Graphics, Inc. Mountain View, CA
- References: <1992Aug28.164008.4015@sics.se> <p6dueak@rhyolite.wpd.sgi.com> <1992Aug30.194536.2545@amd.com>
- Date: Mon, 31 Aug 1992 01:24:53 GMT
- Lines: 49
-
- In article <1992Aug30.194536.2545@amd.com>, shah@angelo.amd.com (Amit Shah) writes:
- > >
- > I have a slight disagreement with your statement that FDDI TTRT by itself
- > does not matter. It depends on the type of traffic you put on the network.
- > If you are talking about bursting traffic in excess of 100 Mbps (such as
- > high-end imaging), with >2 stations on the ring then the
- > TTRT does matter (Lower TTRT will have lower throughput and buffer overflows).
- >
- > However, if you are talking about normal traffic (burst rates not exceeding
- > 100 Mbps) then you are correct within a few percent points.
-
-
- I beg to differ, although I have few kind words for those who force
- customers to have stupidly tiny TTRT's instead of following .
-
- In no case can an old-fashioned, non-Large-Window TCP/IP sender
- transmit more than 64KBytes without allowing the TCP receiver a chance
- to transmit at least one ACK. In other words, the token must be
- released by the TCP transmitter at least once every 5.2 milliseconds.
- The obvious way to send TCP data in steady state is to arrange for one
- ACK per half of the TCP window, or one ACK every 32KB, or one token
- rotation every 2.6 milliseconds.
-
- While the token is rotating, no data is being transmitted. Therefore,
- if your ring latency is large, you might choose to transmit 64KB and
- then wait for one (burst of) ACK(s) every 5.2 ms, in the hope that your
- transmitter's latency in resuming TCP transmissions is less than the
- cost of the dead time. If your systems are fast, this is a reasonable
- hope, since ring latencies can be up to 1.67 ms. If you waste 1 ms on
- token rotations every 2.6 ms, you are not going to get close to the
- speed of light, which is around 97Mbit/sec.
-
- Please notice that 2.6 ms is far lower than the minimum universally
- legal TTRT, and even 5.2 ms is not significantly larger than the silly
- and very wrong (in my opinion) TTRT imposed by default by one large
- vendor.
-
- I agree that a small TTRT with a large or even modest ring latencies
- have bad effects on massive UDP output rates. However, I do not know
- of a realistic application for sending megabytes of UDP data. That
- `ttcp -u` gives such UDP numbers does not by itself make them
- interesting.
-
- These assertions are not based on readings of standards. I haven't
- played with 1 ms ring latencies, but I have spent time with the other
- circumstances I have implicitly described.
-
-
- Vernon Schryver, vjs@sgi.com
-