Application-level versus Network-level Proximity Mohammad Malli PhD student Planete Project, INRIA - Sophia Antipolis France
December 15, Introduction In P2P and CDN, the characterization of the proximity is very important for identifying the best peer to contact or to take as neighbor Our problem is to study how to characterize efficiently this proximity
December 15, Motivation Most of existing protocols rely on the delay for characterizing the proximity of peers Using the delay closest peer as the best peer hides the implicit assumption that the path with the nearest peer has the largest available bandwidth and the smallest loss rate
December 15, Questions Q1. is the delay correlated with other network path parameters ? Q2. does its use approximate the use of the others ? Q3. How much is the performance degradation for applications relying on the delay proximity ?
December 15, Measurement scenario We try to answer these questions using measurements runned over the Planetlab plateform. 127 nodes, spread over the Internet covering America, EU, and Asia, are used. The network characteristics is measured for more than 16k paths.
December 15, Delay vs. available bandwidth 12 % of peers have the maximum available bandwidth on their path with the nearest peer Correlation coefficient between these two parameters is – Slow decrease of the available bw with the delay-based peer rank
December 15, Delay vs. loss rate 88 % of peers have the minimum loss rate on their path with the nearest peer When moving farer in the delay space, the loss rate increases rapidely
December 15, Impact of delay proximity on application performance The weak correlation obtained between the delay and the other path parameters motivates us to evaluate the performance degradation of applications when the proximity is characterized based on the delay alone To this end, we consider two typical applications File transfer over TCP Interactive audio service
December 15, File transfer over TCP We compute the degradation of the transfer time compared to the optimal ranking : File size increases => degradation increases
December 15, Interactive audio service We compute the speech rating factor : Delay proximity => medium audio quality for the 7 closest peers and low quality for the remaining 3 others
December 15, Conclusions Delay is a poor predictor for quality for many applications This can be interpreted by the low correlation obtained between the delay and the other network path parameters Now, we focus on how to deploy the application-level proximity without loosing the scalability
December 15, Thank you Q & A