>        As long as the peer you want to connect to is known by the Metricom
>radios that do wide area forwarding, your peer can be multiple radio hops
>away from you.  If your destination radio isn't known by the wide area
>forwarding service, then you have you have to have a direct connection
>(within radio distance) in order to connect with your peer.  

I'm not sure this is true for all versions of the radio firmware;
starmode-level forarding should also be switchable on the portable
radios if you have suitable access.

In any case, it kind of misses the point that you can do IP-level
forwarding between radio-equipped machines. So you don't __need_
direct connectivity. Though the store-and-forward latency (one extra
hop from each machine to the radio, remember) will get you.

>StarMode, which is what the strip driver uses to communicate with the
>radio, is a packet oriented protocol rather than a stream based protocol,
>you can have multiple peers talk through the same radio.  This is different
>than Metricom's modem-to-modem service where each radio must connect to one
>other radio.

Is the manpage text not clear on that?   How could it be improved?


>        I've used the strip driver, driving two radios through the wide area
>forwarding service under NetBSD 1.2, and it worked great.  Too bad the
>radios weren't't faster.  The long round trip delay drove TCP performance
>into the ground.

Uh, the newer radio firmware significantly improves latency on the
serial link between computer and modem.  But the real latency killer
is when you forward through multiple radio hops: the radios are
half-duplex and the frequency-hopping means you can only send and
receive about 10 packets/sec.  And if you have bad reception the
radios go into a hop-by-hop error-correcting mode, which nearly
doubles per-send time.  Other than that, TCP handles throughput pretty
well. It's latency that loses. Telnet over a single radio hop can
become quite tolerable.


Stuart Cheshire's Hot Interconnects 96(?) paper has more details, and
there should be lots of gory graphs on the MosquitoNet.Stanford.EDU
Web site.  Caveat: many of the numbers there were measured on Linux
systems. Linux's quaint and picturesque networking stack did not do
Nagle SWS avoidance before about Linux 2.0.29.  Those numbers may
not be representative of BSD TCP performance over Metricom.