CURRENT_MEETING_REPORT_
Reported by Osmund de Souza/AT&T
OPSTAT Minutes
The proposed Agenda for the meeting was:
o Administration
o Metrics - old, new, and exotic
o Measurement polling periods
o Query language and exchange protocol
o Report formats
o Closing
The proposed Agenda was approved. Osmund de Souza volunteered to take
the Minutes.
Metrics: The group reviewed the metrics that have been proposed so far.
Bernhard had written a ``metrics paper'' before the meeting to help
focus the discussion. We agreed that the basic set of measurements
should include the following:
o Octets in/out, unicast packets in/out, non-unicast packets in/out
for each interface.
o IP packets forwarded, IP packet discarded, (similar counts for
other network layer protocols), for each router.
We then talked about whether we should expand this set to include new
and exotic metrics and if so, what they should be. There were ideas
about measuring availability, stability, delay performance, congestion,
and line errors. We decided that rather than try to come to a
conclusion at the meeting we would take it to the mailing list. We did
agree that the metrics paper should not explicitly include or exclude
these new measures, but rather should leave the door open for expanding
the base set later. We also agreed that initially it would be wise to
work with variables within the current MIB structure, though this should
not limit us as the opstat architecture matures.
Measurement: As usual, we had a lively discussion about measurement
polling intervals. Vikas Aggarwal presented the results of a study he
did on JvNCNet. In the study he periodically polled each interface in
the network to download traffic data to a central manager. Since it
took three minutes to poll all the interfaces, the polling period was
three minutes. He computed the utilization of each link in the network
for each three minute interval, and also for integer multiples of three
minutes by aggregating the measurements. The results showed that as the
``integration'' interval increased, the average of the average
1
utilization in each interval remained unchanged (to within some small
error) but the variation in average utilization for each interval
decreased. Hence, for instance, the three minute utilizations had
several peaks above 70
The discussion led to the understanding that measurement collection
should ideally be done on a small enough time scale to capture
short-term variations and peaks in traffic loads. Burstier traffic
requires a smaller polling interval. However, small interval polling on
a large network may be difficult for most network management systems,
and there is the problem of storing the measurements for later
processing. While no definite conclusion was reached on this one, the
group is close to settling the issue by gravitating towards a polling
interval of 5-15 minutes.
Query Language: We reviewed the paper that Bernhard had put together
(and circulated on the mailing list) to define a query language syntax
for the statistics client/server model. The language is geared around
the following simple commands: login, exit, help, format, and select.
There was some discussion about the meaning of the ``format'' command
and what help the ``help'' command was supposed to give. We decided
that ``format'' should let the user know the storage/presentation format
of the data in the query, and that ``help'' should explain the meaning
and syntax of commands as well as the nature of the data named in the
query. Bernhard agreed to incorporate our comments in the document
before resending it to the list.
David O'Leary said that he may be able to get a (more?) graduate
student to work on a statistics client/server protocol. He did not have
a clear idea of the direction in which the work would proceed, but would
keep the group updated through the mailing list.
Report Formats: We did not have enough time to spend discussing the
format of the reports. We reviewed the ideas that had come up at
previous meetings and that Bernhard had summarized in a ``reports
paper''. We did agree that we may have spent too much effort in the
past defining the ``reports for upper management'', aka The McDonald's
Report. We felt that we should concentrate on engineering reports for
network operations and planning. These reports could then be stripped
of an appropriate amount of useful information to present to upper
management.
Closing: An item that came up during the discussion was the need for us
to interface more with other groups, notably SNMP, and Internet
Accounting. We were also pleased about the progress we had made since
our previous meeting in St. Louis.
Finally the meeting agreed to use the papers produced by Bernhard as a
basis for the continued work.
Attendees
Thomas Brisco brisco@rutgers.edu
Gigi Chu gigic@hpspd.spd.hp.com
2
Henry Clark henryc@oar.net
Shari Galitzer shari@gateway.mitre.org
Kenneth Goodwin goodwin@psc.edu
Phillip Gross pgross@nis.ans.net
Michael Khalandovsky mlk@ftp.com
Tim Lee-Thorp ngc!tim@uunet.uu.net
Chris Myers chris@wugate.wustl.edu
David O'Leary oleary@sura.net
Robert Reschly reschly@brl.mil
Mark Saake saake@llnl.gov
Timothy Salo tjs@msc.edu
Erik Sherk sherk@nmc.cit.cornell.edu
Frank Solensky solensky@clearpoint.com
Bernhard Stockman boss@sunet.se
Roxanne Streeter streeter@nsipo.nasa.gov
David Waitzman djw@bbn.com
Chris Waters-Pierandozzi waters@jvnc.net
Gerard White ger@concord.com
Osmund de Souza desouza@osdpc.ho.att.com
3