Re: Last Call: Security Architecture for the Internet Protocol to Proposed Standard

[Phil Karn <karn at qualcomm.com>]

The TCP-over-satellite group is misnamed. It should be called the
TCP-over-large-bandwidth-delay-paths group. Many terrestrial paths now
have bandwidth*delay products as large (or larger) than satellite
paths.

Cable modems are an increasingly important example. One very popular
modem, the Motorola CyberSURFR, has a 27 Mb/s downstream path (limited
to 10 Mb/s by the 10-Base-T interface) but only a 768 kb/s polled
upstream path. Delays in the upstream direction are highly variable,
frequently reaching several hundred milliseconds under evening
load. (Contributing factors: other users doing large uploads and a
cable company that's notably stingy about adding resources).

Analyze a download over such a modem and you can see how large
bandwidth*delay products are not confined to satellite paths. Worse,
most of these cable modems support only one directly-connected host
computer running the Windows 95 TCP/IP stack. This has spawned a minor
cottage industry in programs to increase the TCP receive window size
in the W95 system registry. But the W95 stack doesn't support window
scaling, so 64K is the limit. The only solution to the problem is to
get Microsoft to update their stack, as difficult as that may be.

(Suggestions that the cable modems should silently proxy every TCP
connection will be greeted with derisive silence.)

--Phil