Providing safety and visibility for mobile users.

Abstract

Mobile users bring new challenges to distributed file systems. First, the network costs between clients and servers vary due to mobility of clients. Second, mobile clients are less available due to the absence of network connectivity or being suspended for power savings. Third, mobile clients are unreliable because they are highly susceptible to breakage and theft. Current file systems needlessly combine safety and visibility; a client propagates the contents of an update, implicitly notifying the server that the update exists. Fluid Replication separates them through the addition intermediate servers, called WayStations. While traveling, a client associates itself with a nearby WayStation that provides replication services. Updates are sent to the nearby WayStation for safety, while WayStations and servers frequently exchange knowledge of updates through reconciliation to provide visibility. In this way, Fluid Replication provides safety and visibility to file updates over the wide-area with performance comparable to the local-area. To choose a nearby WayStation, clients need to know the available network capacity to WayStations. This is particularly difficult in mobile networks where capacity changes frequently. Current systems depend on static exponentially weighted moving average (EWMA) filters. These filters are either able to detect true changes quickly or to mask transients, but cannot do both. This motivated the design of a new network filter. This filter is agile when possible and stable when necessary; it adapts to the prevailing network conditions. During reconciliation, a WayStation sends the server update notifications without the contents of those updates. While sending notifications frequently is important to improve consistency, sending data should be deferred to reduce server load. However, deferring of data shipment penalizes clients who share because to serve these clients, the server must fetch updated files from other WayStations, called back fetch, over a wide-area network. To solve this problem, a heuristic that predicts future file sharing has been developed. It is based on the observation that past instances of sharing are likely to lead to future ones. It reduces data shipped by orders of magnitude compared to aggressive schemes, while reducing the number of back-fetches by nearly half compared to on-demand shipment.