problems using afsfsperf to measure AFS performance to remotecells

[Harald Barth]

> I'm not sure I'm going to be able to delve this deeply into this until
> after I've submitted my PhD thesis (the main motivation for performance
> tests) which is now more or less imminent!  However, I'll be interested to
> pick this up again, possibly mid-April.

> ... ARLA works fine, it's just the afsfsperf that seems to fails ...

I may have a not-so-deep-suggestion: One of the few (only?) things
that differs between rx running from arlad and rx running from
afsfsperf is the UDP port:

habarber:perf$ grep 471 afsfsperf.c 
    ret = rx_Init (htons(4712));

Arlad on the other hand uses 7001, the same as OpenAFS. You can test if
you can access the other cells with arla-cli, it uses 4712, too. If you 
have the same problems with arla-cli, you may want to find the guilty
firewall or change this in the source. If this cutting down of ports
continues in this rate, in some future only port 80 and 443 will be left
to communicate on the Internet. Maybe 20 and 21 because we have allways
and will allways use ftp ;-).

Why 4712? Because arlad was using 4711. Why was arlad using 4711 then?
4711 is on of the two well know "random number" (used in a lot of
examples) at KTH. The other one is 17. For example:
http://www.sgr.nada.kth.se/misc/tex/nadabrev/exempelbrev.pdf

Please let us know how your perftests work out. It will all be in your
PhD thesis, well? My personal guess is that rx today suffers from
window size problems which make it latency bound.

Harald.