Speaker: Patrice Abry, CNRS, Research director ENS Lyon, France (joint works with P. Borgnat, G. Dewaele, D. Veitch, N. Hohn, K. Fukuda, K. Cho, F. Ricciato, R. Fontugne)
Date: Friday, January 31, 2014
Time: from 15:00 – 16:00
Place: National Institute of Informatics, 20th floor, room #2001
Prof. M. Taqqu was a member of the team that first evidenced long memory and scaling in Ethernet traffic in the mid 90ies. He was then at the origin of a celebrated theorem that established a theoretical connection between heavy tails and long memory mimicking classical queuing and service mechanism in
networks. These seminal contributions paved the way toward numerous analysis of heavy tails, long memory and scale invariance at large in aggregated Internet traffic time series. Despite signi ficant efforts,
Taqqu’s theorem has been found difficult to verify on real aggregated Internet traffic time series. In this talk, we illustrate that these difficulties result both from the asymptotic nature of the theorem and the need
for robust estimation. Robust estimation is achieved by combining median and random projections (or sketches) and illustrated at work on 7 years in the life of Internet (MAWI database, transpacifi c US to Japan traffic). While heavy tails explains long memory at coarse scales (above the second), scaling at fi ne scales (below the second) are also discussed in terms of models (Cluster Point Process versus Multifractal) and applications (anomaly detection).