Some Thoughts for Position Papers (Getting the Creative Juices Flowing)

NSF Internet Statistics Measurement and Analysis Workshop February 1996, San Diego.

Below I've written some rough notes on topics which may be covered in the workshop. There may be other things which I neglected, and of course, we will not cover all of this.

Invitees are asked to write short statements representing the interests and/or opinions of their company or organization. Specifically, you should describe the topics that you consider to be most important as items for discussion at the workshop. We hope to identify important problems and questions for the Internet industry as a whole and either achieve actual consensus on solutions or initiate a process which can solve the problems we identify.

-Mark Garrett

General:
What do various players care about, what motivates them (identified by position papers)? Focus is the 1-3 year time frame, not many-years-out research. What can be done with current technology with minor modifications?

Policy:
What measurements of Internet traffic statistics would an Internet Service Provider (ISP) be willing to disclose in order to facilitate development of traffic management tools? How could such measurements be filtered to release only non-sensitive or customer-confidential information?

Industrial cooperation.
Importance of cooperation in absence of any central control. Role of NSF, ARPA and other government agencies going forward. Discuss barriers preventing interested parties from doing what they might; actions to take to remove barriers. Discuss what might happen when obstacles are removed.

Legal Issues.
Implications of measurement, privacy etc. Privacy of ISP and of their customers' information (e.g. TCP port numbers).

Network dimensioning
The obvious questions about how big to make bandwidth and buffer resources, given some knowledge of the number and types of connections/applications etc. What measurement can be used to predict capacity shortage? Average load metrics are insufficient. Provider needs to be efficient while not losing packets.

Traffic Modeling.
Meaningful characterization of bursty nature of traffic. Average load or rate measures don't predict resource consumption consistently. Models based on real measurements (e.g. Long-range dependence).

How can the Internet community leverage the large set of academic experts working on teletraffic models, and get them focused on real-world problems with realistic assumptions etc? (The solutions may not be so powerful and elegant mathematically, but they will impact real networks.)

Current Traffic Measurement.
What is going on now to measure statistics and performance in today's Internet (e.g. mtrace is used with MBONE traffic to find loss points)?

Network integrity.
How does network provider track down faulty equipment and isolate sources of loss when user complains. Relation between network provider and customer when much of the Internet traffic control mechanism resides at the end points. What background measurements and alarms need to be in place?

Internet network traffic analysis.
Advance the state of network analysis in general. How to tell that this is happening. (One of NSF's main interests for the meeting.)

What should NSF and other agencies do to facilitate the advancement of statistics collection that will lead to improved understanding of network behavior?

How should NSF foster collaboration on these topics in a competitively supplied environment?

How can NSF use the vBNS and NAPs to prototype and disseminate developments in wide area statistics collection/monitoring? Should a more rigorous operational statistics architecture be explicitly specified? Should the NAPs be required to develop their own or reach consensus on a common model, perhaps in conjunction with the Routing Arbiter? Is there a particular forum (EOWG, IEPG, Farnet) in which these communities should make regular presentations on behalf of NSF?

Switch dimensioning.
What are the implications of real measurements of traffic on switch design (size of tables, call set-up capacity, number of concurrent calls, set-up latency, holding time, etc.).

Fault Prediction.
Probability model for prediction of network fault occurrences. What role for statistics measurement?

Security.
Can data be measured related to e.g. spoofing attacks, etc. Do people really have hard data evidence? Are people interested in seeing such data? Knowing what is going on the lines for security ends.