System evaluation

This section evaluates the design and implementation of Pangaea. First, we investigate the baseline performance and overheads of Pangaea and show that it performs competitively with other distributed file systems, even in a LAN. Further, we measure the latency, network economy, and availability of Pangaea in a wide-area networking environment in the following ways:

• We study the latency of Pangaea using two workloads: a personal workload (Andrew benchmark) and a BBS-like workload involving extensive data sharing. For the personal workload, we show that the user sees only local access latency on a node connected to a slow network and that roaming users can benefit by fetching their personal data from nearby sources. Using the second workload, we show that as a file is shared by more users, Pangaea progressively lowers the access latency by transferring data between nearby clients.

• We demonstrate network economy by studying how updates are propagated for widely shared files. We show that Pangaea transfers data predominantly over fast links.

• To demonstrate the effect of pervasive replication on the availability of the system, we analyze traces from a file server and show that Pangaea disturbs users far less than traditional replication policies.