IP Performance Metrics Working Group

Initially a summary of the work that has already been done was presented, this included the following:

• Path Performance
• Path Availability

Work should be done on the Multicast Area.

There must be two general methodologies for measuring a certain metric.

Vern Paxson presented a revised framework document for the working group activities. Then Guy Almes talked about 1-way delay measurement and Packet Loss measurement.

The statistics for both metrics will involve:

• minimum
• percentile
• median
• inverse percentile

Typical values for Packet loss should be less than 1 the measurements should be a Poisson process.

There is an ongoing activity by a collaboration of 23 U.S Universities involved in the definition of performance metrics for IP. Among them the Universities of Chicago, Colorado and Michigan are fully meshed. True Time GPS clock synchronisation is being used which provides an accuracy of 10 microseconds.

There was also an argument as to when the arrival time or the send time should be calculated at either the Application level or at the Kernel level.

It was also stated that the delay characteristics vary in each direction and the reasons for this should be further investigated especially when the assymetric path case is not true.

Following was a presentation by the IPMA Project (www.merit.edu). The NetNow daemon and results of monitoring for BGP pathologies.

In the list of things to be done Vern Paxson specified:

• tests on correlation
• definition of clouds as directed graphs
• broadcast networks to be defined as a set of links

Macroscopic Behaviour of TCP Congestion Avoidance, paper published in July's Computer Communication Review, work by the Pittsburgh SuperComputer Centre

Performance Measures for Voice over IP from NOKIA. The metrics of interest here are :

• end to end delay
• jitter
• packet loss