> Does anyone have any kind of suite of tools set up to do performance
> monitoring? I just want some simple things, like how many bytes/second
> is the server putting out files, how long does it take to download files
> to a client (Again, bytes/second), etc...
Hi,
Well, this is my current thesis effort actually! I haven't
found a lot of material, but I've been developing some of my own tools
and I'll tell you what I have so far. By the way, what use do you plan
to make of these statistics, and what is your time scale of interest?
There are a lot of loose variables that might cause your data to vary
wildly, since (statement of the obvious coming) AFS is a multi-part
system running on many types of clients and servers with different
CPU's, OS's, and storage technologies, and existing on a set of
networks used concurrently for other services. (My plan of attack is
to use large time spans and all our servers, to let the law of
averages let the data settle down.)
Here we go:
http://www.citi.umich.edu/u/honey/papers.html - Peter Honeyman has
done a lot of measurement with AFS. Check out "AFS Server logging" and
othe rinteresting papers on this page.
afsmonitor (for which I posted a Solaris patch earlier; did Transarc
ever update this?) - this is an unweildy dog that can nevertheless
continuously monitor various useful statistics, such as FetchData and
StoreData on a per-fileserver basis, in KB and # of actual requests.
scout - a smaller program with similar abilities, uses the same query
protocol
the xstat package - the Rx library that sets up monitoring threads for
you, upon which the above tools are based. Header files are libraries
are availabe in the standard distribution.
querystats - an abstraction-thumping kludge I put together that has
the features of afsmonitor I want (monitoring of an arbitrary xstat
statistic) without the problems I don't (bound to fixed port on client
(?!!?), inflexible display and statistic storage options, no C
interface to the statistics, inability to precisely time and readjust
query intervals)...well, it works dependably. I'm not prepared to
release it on the world, but I will work together with any interested
parties.
Other options are to run tcpdump or a sniffer on the AFS server
subnetworks, and write your own scripts to collect per-port
statistics. (Although it's a lot of extra work at this level to
distinguish individual server-to-client conversations.)
I'm still looking for more information, insight, and methodology as
well. I'd like to be a repository for AFS statistics and performance
information, since it is my current area of focused research. If
everyone would e-mail whatever useful things they can tell me I can
digest and disseminate the information perhaps via the Web.
Thanks,
Daniel Bromberg, Co-op
[EMAIL PROTECTED]
M/S 171-300 (818) 393-3872
Jet Propulsion Laboratory
4800 Oak Grove Dr.
Pasadena, CA 91109
