Description
Replicated systems often use quorums in order to increase their performance and availability. In such systems, a client typically accesses a quorum of the servers in order to perform an update. In this paper, we study the running time of quorum-based distributed systems over the Internet. We experiment with more than thirty servers at geographically dispersed locations; we evaluate two different approaches for defining quorums. We study how the number of servers probed by a client impacts performance and availability. We also examine the extent to which cross-correlated message loss affects the ability to predict running times accurately from end-to-end traces.