Re: Packet TCP/IP handled by the OS

[glg@balrog.k8lt.ampr.org (Gary L. Grebus)]

>I read some (old) discussion of this in the tcp-group archives, from earlier
>this year.  If the intent is to handle most of the communications at the
>OS level, and less by a "specialized" program like NOS, are the OS's going
>to have to be drastically modified to handle this?  Or is a parametric "tune-
>up" enough to make it work?

I've been running a FreeBSD system hooked to the local 1200 baud
network (via a NOS router).  To get it to work as well as a NOS
system I did some some tuning and some source code tweaks:

1. Associate an MSS and IRTT value with each statically configured
   route.  I think most systems with recent Berkeley-derived network
   code have this as a feature.  This is even better than what you can
   do with NOS, which only allows one default irtt and a per-interface MSS.

2. Changed the tcp timer code to continue to retry once the maximum number
   of retransmissions is reached, rather than closing the connection.
   This is consistent with the way most NOS systems are run around here.

3. Lengthened the initial value of the timer used to time connection
   establishment.  This lets you deal with really long rtt's.

4. Changed the table that controls retransmission backoff to start out
   more linear before switching to an exponential curve.  Yeah, I
   know this is unclean.  But packets do get dropped more often over
   the radio, even without congestion.

5. Wrote a utility that does NOS-like "tcp stat" and "tcp kick" commands.

(Anyone who wants the patches or utility is welcome.  Drop me some mail.)

Of course, with  faster, more reliable links some of these issues
may not be important.

At one point in the past, I used some of the same tweaks by making binary
patches to a commercial UNIX system, but it's sure much easier with
source code.

This topic came up on the nos-bbs list a couple of weeks ago.  The
main point I made there was that it is possible to use existing stacks
(Winsock, OS/2, or UNIX) over radio, but that it will be a whole lot
easier (and work better) if we can modify the code to fit the local radio
environment.

73,
        /gary

Gary L. Grebus          Home:  glg@k8lt.ampr.org
Brookline, NH           Work:  grebus@eng.pko.dec.com
                        Ham Packet: K8LT@WA1PHY.#EMA.MA.USA