Topics: Network analysis and design; Optimization techniques; Traffic engineering
Authors: Gerhard Hasslinger (Deutsche Telekom, Germany); Kostas Ntougias (Athens Information Technology, Germany); Frank Hasslinger (TUDa, Germany); Oliver Hohlfeld (RWTH Aachen University, Germany)
Presenter bio:
Gerhard Hasslinger has 10 years of experience as a researcher and
lecturer in computer science at TU Darmstadt, as well as 20 years of
practice in engineering of fixed/mobile broadband IP networks and
services at Deutsche Telekom, Darmstadt, Germany. His research interests
include content distribution, traffic engineering, reliability, and
quality of service aspects of computer and communication networks, as
well as information theory and coding.
Abstract: The topic of Internet content caching regained relevance over the last
years due to extending data center infrastructures in CDNs, clouds and
ISP networks to meet capacity and delay demands of multimedia services.
In this study, we evaluate the performance of web caching strategies in
terms of the achievable hit rate for realistic scenarios of large user
populations. We focus on a class of score gated least recently used
(SG-LRU) strategies which combine the simple update effort of the usual
LRU policy with the flexibility to keep most important content in the
cache according to a predefined score function. Caching efficiency is
evaluated via simulations assuming Zipf request pattern, which has been
confirmed manifold on popular web platforms for video streaming and
other types of content. For the standard independent request model
(IRM), we analyze the possible hit rate gain of other web caching
strategies over pure LRU within the complete relevant range of three
basic system parameters. The results confirm that absolute hit rate
gains of 10%-20% over LRU being observed in some case studies for
alternative caching strategies are not optimistic but a realistic
estimation in general. Finally, we compare IRM evaluations with results
for varying request pattern over time based on Wikipedia statistics,
which recently have been made available for daily top-1000 page
requests. Simulations are extended to show the impact of varying object
popularity on the caching efficiency and to optimize the caching
strategy with regard to increasing popularity dynamics.
