Estimating network statistics from distributed packet streams in high speed networks

• 作 者: Jin Cao
• 机 构: BellLabs,Alcatel-Lucent,MurrayHill,NJ,USA
• 出 处: IMS-China International Conference on Statistics and Probabi
• 发表日期: 2008年6期

　　Many aspects of network management, such as network provisioning, traffic en gineering, and traffic anomaly detection, rely on knowing the traffic flow statistics between origin and (lestination (OD) node pairs, However, rapid growth of networks in size and bandwidth makes obtaining accurate OD traffic flow statistics very chal lenging.In this talk, we addrcthe the problem of OD flow and byte counting from the point view of streaming data analysis.Our mcthod constructs a compact sketch via hashing and randomization for each data strcam, and derives efficient estimates of the OD counts, ising the pair of sketches received at a central location.We show how our method can be generalized tbr counting unique flows across more than two distributed data streans, and derive accurate approximation of the error distribu tion of our estimates.Wc demonstrate our algorithms yields much more accuratc estimates than existing methods including network tomography.