Network Traffic (George Michalidis , organizer)

Andre Broido (Caida)
Diversity and Disparity in IP Traffic

Friday 8:30-9:00, Fountain II

Abstract:

The need to service populations of high diversity in the face of high disparity affects all aspects of network operation: planning, routing, engineering, security, and accounting. Mathematically, disparity is a property of two almost disjoint measures- a counting measure and a size measure- on the same set of objects. We define crossover disparity as the fraction c of total volume contributed by a complementary fraction (1-c) of large objects- e.g. 80% of traffic from 20% of sources. We evaluate this metric for common distributions such as Pareto. We then use it to define a boundary between mice and elephants in traffic aggregated by IP addresses, prefixes and ASes, Studying sources and sinks at two Tier 1 backbones and one university, we find that 90:10 and 95:5 are common crossover values for IP traffic. We then discuss the origins of disparity in business and operational realities of IP networks, and its mathematical underpinnings. Our results will be useful for developers of traffic models, generators and simulators, for router testers and network operators. This is a joint work with Young Hyun, Ruomei Gao, and K.C. Claffy.


George Michailidis (University of Michigan)
Flexicast Delay Network Tomography

Friday 9:00-9:30, Fountain II

Abstract:

Characterizing network performance is an important problem for service providers. In this talk, we discuss link delay assessment based on active network tomography; i.e. characterizing internal link delay distributions based on path delays of injected traffic. We consider a new class of probing experiments, called flexicast, that allows selective investigation of the network. We examine several estimation schemes and the associated identifiability issues. The methods are applied to a VoIP readiness study.


David Rolls (UNC Wilmington)
Semi-experiment Investigations of Network Traffic

Friday 9:30-10:00, Fountain II

Abstract:

TCP/IP flow data involves a hierarchy of details. For example, at the flow level there are flow starts and durations. Within a flow there is the pattern of flow times. The semi-experimental method involves selectively altering one or more of these factors, leaving the rest unchanged. We combine semi-experiments with queueing metrics and statistics to understand how the factors contribute to the traffic's queueing profile. This is joint work with Ericson Davis and George Michailidis.