Check out the new USENIX Web site.

Performance Analysis

In Section 1 we identified two main reasons for data replication on the Web: better reliability and better performance. The reliability improvement is obvious: If a single server has a mean time to failure MTTF1 and mean time to repair MTTR1, then a system with n fully replicated servers has a mean time to failure given by

\begin{displaymath}MTTF_n \approx \frac{MTTF_1^n}{n \cdot MTTR_1^{n-1}}

assuming that the server failures are statistically independent [21]. Clearly, the mean time to failure improves with the number of replica servers. In order to ascertain the performance gains of resource replication, we conducted a live experiment with the Web++ system.

Figure 9: Experimental configuration.