problems using afsfsperf to measure AFS performance to remotecells

[Harald Barth]

> I've managed to install ARLA on a Fedora Core 1 Linux PC 
> (2.4.22-1.2197.nptl stock kernel) along with kth-krb.  I want to use the 
> afsfsperf program, as one perspective, to look at the performance of AFS 
> over long distance networks...

Congratulations.

> It also has no problem with a cell 250 km (ping: RTT ~= 5.5ms) away: 
> rl.ac.uk (although I had to use klog from OpenAFS to get a 
> token, rl.ac.uk didn't like kth-krb kinit from here?!!)

What program you need depends what authentification the cell admins
have decided upon. I'd with the world would see that Krb5 is the way
to go and throw away old krb4 and kaservers.

> When I try to run afsfsperf to desy.de (RTT~=34ms) I get the following
> $ /usr/arla/sbin/afsfsperf -host vayu.desy.de -volume 536903044 -authlevel 3 -cell desy.de -smallrun
> afs low-level fileserver performance tester
> afsfsperf (arla 0.36.2)
> Copyright 1999-2003 Kungliga Tekniska Hgskolan
> Copyright 2002-2003 Stockholms universitet
> Send bug-reports to arla-drinkers at stacken.kth.se
> Perf
> host: vayu.desy.de, volume: 536903044, auth: crypt
> afsfsperf: RXAFS_GetTime returned 4801

I don't _know_ what's wrong, but I can give you some leads to follow.

1) Arla's rx is similar but not the same as OpenAFS's rx.

2) OpenAFS's rx _may_ fragment rx packets (search for RX_MAX_FRAGS in
   the OpenAFS source). This will hit you if you have a router in the
   path that drops fragmented packets or if you have very high packet
   loss. However, small rx packets like GetTime should get through.

3) Could it be that you have problems with authentification?

4) To proceed further, you probably need to look at the rx versions
   (rxdebug -version) at the remote sites and use tcpdump and ethereal
   (modern ethereal has quite good rx support) to look what's going
   on.

Work to use TCP has started, but because of the "only 24hours/day issue"
not getten further than the specification. 

/afs/stacken.kth.se/projects/arla/hackathon-2004/tcp/rxtcp

Unfortunately, noone seems to want performance in AFS desperately
enough to actually fund some work. Instead a lot of resources are
thrown into other pits like NFS (insert rant).

On the other hand someone (hey, don't you guys look all at me) should
compile a afsfsperf against OpenAFS's rx and see if there is any
difference. And probably tell the rest of the world how to do that
with the least amount of hassle.

Harald.