SANE: Stable Agile Network Estimation

Distributed systems are becoming increasingly dependent on network estimation, the ability to determine performance along one or more network paths. Producing quality estimates is challenging because network observations are noisy; this is particularly true in wide-area or mobile settings. Current systems depend on simple exponentially weighted moving average filters. These filters are either able to detect true changes quickly or to mask transient changes, but cannot do both. In this paper, we present four filters designed to react quickly to persistent changes while tolerating transient ones. These filters are evaluated in a variety of networking scenarios through three metrics: agility, stability and accuracy. While no single filter dominates, one based on techniques from statistical process control shows promise relative to the others.