[290] in Info-AFS_Redistribution
Re: Cross cell monitoring survey.
daemon@ATHENA.MIT.EDU (Michael T. Stolarchuk)
Mon Aug 19 18:08:58 1991
To: Steve Mattson <hobbes@caen.engin.umich.edu>
Cc: Info-AFS@transarc.com, afs_systems@caen.engin.umich.edu,
In-Reply-To: Your message of "Fri, 16 Aug 91 13:37:58 EDT."
Date: Mon, 19 Aug 91 12:11:46 -0400
From: "Michael T. Stolarchuk" <mts@terminator.cc.umich.edu>
We use a tool here, which listens to promiscuous ethernet
traffic (using the enet filter, or nit), and then breaks
out RX packets into its details.
One version of the tool records the start time and end time
of the different AFS ops, for each of the servers and for
the 'total' of the servers (along with std deviation).
I didn't say that very well... it makes pages... each page
describes the duration, number, avg, and std dev for some
grouping of the AFS operations recorded.
One of the pages is 'all AFS ops'. When we first made the
tool, the std dev was huge compared to the avg of some of the
ops. We then make 'pages' for each of the fileservers, and
found the large std dev was partially due to the different
service times of the machines we see.
The tool is really helpful to determine if there is a
server which is relativly slow compared to others. Its
helpful too, since it shows all the servers the local
AFS clients are using; not just the local cell, other
cells too. And it sees the service time from the clients
point of view, not the servers. We have instrumentation
for file servers (as transarc does), and the values recorded
by the servers and the 'this tool' are a bit different.
mts.
ifs, uofm
