Modeling of Dynamic Latency Variations for Simulation Study of Large-scale Distributed Network Systems

A realistic evaluation model that simulates a realistic variation in latency between a large number of Internet nodes is attracting attention for use in simulation studies on large-scale distributed networks (e.g., CDN & on-line gaming). Therefore, we are attempting to build a new evaluation model by applying time series analysis using an ARIMA model and a Euclidean embedding technique to a dataset obtained from a global scale measurement service for this study. The proposed evaluation model not only generates accurate time series data of the latency of each path on the Internet, but also avoids unrealistic behaviors related to the spatial distribution of latency (i.e., triangle inequality violation). Furthermore, the performance of a representative distributed latency management and prediction system is evaluated through computer simulations as a use case of the proposed evaluation model. The evaluation results helped us to clarify whether or not the proposed model can correctly disclose the impact of the dynamic latency variation on large-scale distributed systems. An example program of the proposed evaluation model is found in an Appendix of this manuscript.