An Empirical Study of Bandwidth Predictability in Mobile Computing

While bandwidth predictability has been well studied in static environments, it remains largely unexplored in the context of mobile computing. To gain a deeper understand-ing of this important issue in the mobile environment, we conducted an eight-month measurement study consisting of 71 repeated trips along a 23Km route in Sydney under typ-ical driving conditions. To account for the network diver-sity, we measure bandwidth from two independent cellular providers implementing the popular High-Speed Downlink Packet Access (HSDPA) technology in two different peak access rates (1.8 and 3.6 Mbps). Interestingly, we observe no significant correlation between the bandwidth signals at different points in time within a given trip. This observa-tion eventually leads to the revelation that the popular time series models, e.g. the Autoregressive and Moving Average, typically used to predict network traffic in static environ-ments are not as effective in capturing the regularity in mo-bile bandwidth. Although the bandwidth signal in a given trip appears as a random white noise, we are able to detect the existence of patterns by analyzing the distribution of the bandwidth observed during the repeated trips. We quantify the bandwidth predictability reflected by these patterns us-ing tools from information theory, entropy in particular. The entropy analysis reveals that the bandwidth uncertainty may reduce by as much as 46 % when observations from past trips are accounted for. We further demonstrate that the band-width in mobile computing appears more predictable when location is used as a context. All these observations are con-sistent across multiple independent providers offering differ-ent data transfer rates using possibly different networking hardware