Affiliations: National Key Laboratory of Networking and Switching, Beijing University of Posts & Telecommunications (BUPT), Beijing, 100876, P.R. China E‐mail: linyu@bupt.edu.cn, chsd@bupt.edu.cn, cgwang833@yahoo.com, xiaodan@bupt.edu.cn, zoush@bupt.edu.cn | Faculty of Communication and Information Engineering, ChongQing University of Posts & Telecommunications (CQUPT), ChongQing, 400065, P.R. China E‐mail: longkp@cqupt.edu.cn
Note: [] Corresponding author.
Abstract: The p‐pair/train technique estimates the capacity of a path from the dispersion experienced by two/multiple back‐to‐back packets. The dispersion of packets may follow a complicated multimodal distribution, which causes statistical filtering methodologies and heuristic approaches (e.g., pathrate) to be inaccurate in heavily‐loaded cases. This paper analyzes the noise characteristics in the path capacity measurement, and elaborates on the distribution of capacity estimates. Based on this analysis, a new algorithm called LMSA (Local Modes Statistical Algorithm), and its enhancement (LMSA+) are presented. Using multiple distributions of capacity estimates obtained by variable‐sized probing p‐pairs, LMSA+ only picks out the local modes that locates at fixed positions, and selects the minimum of these modes as the path capacity. Simulation results demonstrate that, compared with pathrate, LMSA+ is more accurate in heavily‐loaded paths or paths with relatively large bottleneck bandwidth values, and is more robust to the dynamic cross traffic load.
Keywords: Active network measurements, bandwidth monitoring, bottleneck bandwidth, available bandwidth, measurement noise
