Medium-term Client-Perceived Performance Prediction

摘要

—In recent years the networking infrastructure has improved and as a result there has been a tremendous growth in the number of providers and the services they offer. With the wide choice of available services, many clients are interested in differentiating providers based on Quality of Service (QoS) - performance being one of the most important QoS attributes. In this paper we focus on provider differentiation based on client-perceived performance. The client-perceived performance better represents client experience compared to server-side performance measurement used widely today. We analyze and characterize client-perceived performance, based on Internet measurements. Based on this characterization, we propose a technique - last-period prediction (LPP) - for medium- term performance prediction between a client-provider pair, which will be used for provider differentiation. The LPP technique is specially designed to capture the characteristics of client-perceived performance, such as periodic fluctuations during the concerned period. Through experiments on the Internet, we demonstrate that the proposed technique provides better prediction accuracy for medium-term prediction of client-perceived performance compared to some popular time series models such as ARIMA, seasonal ARIMA, exponential smoothing, and Holt-Winters.