Safety, Visibility, and Performance in a Wide-Area File System
Minkyong Kim, Landon P. Cox, Brian D. Noble · 2002
As mobile clients travel, their costs to reach home filing ser-vices change, with serious performance implications. Current file systems mask these performance problems by reducing the safety of updates, their visibility, or both. This is the re-sult of combining the propagation and notification of updates from clients to servers. Fluid Replication separates these mechanisms. Client updates are shipped to nearby replicas, called WayStations, rather than remote servers, providing inexpensive safety. WayStations and servers periodically exchange knowledge of updates through reconciliation, providing a tight bound on the time until up-dates are visible. Reconciliation is non-blocking, and update contents are not propagated immediately; propagation is de-ferred to take advantage of the low incidence of sharing in file systems. Our measurements of a Fluid Replication prototype show that update performance is completely independent of wide-area networking costs, at the expense of increased sharing costs. This places the costs of sharing on those who require it, pre-serving common case performance. Furthermore, the benefits of independent update outweigh the costs of sharing for a work-load with substantial sharing. A trace-based simulation shows that a modest reconciliation interval of 15 seconds can elimi-nate 98 % of all stale accesses. Furthermore, our traced clients could collectively expect availability of five nines, even with deferred propagation of updates. 1