next up previous
Next: Summary Up: Evaluation Previous: Flash Crowds

Zipf Parameter Change

Finally, we examine the adaptation of Beehive to global changes in the parameter of the overall query distribution. We issue queries from Zipf-like distributions generated with different values of the Zipf parameter at each 24 hour interval. We seeded these simulations with 4096 objects.

Figure 8 shows the lookup performance of Beehive as it adapts to changes in the parameter of the query distribution. Beehive effectively detects changes in the Zipf parameter and redistributes object replicas to meet targeted performance objectives. Figure 9 shows the average number of objects replicated at each node in the system by Beehive. As the Zipf parameter varies, so do the number of replicas Beehive creates on each node. Beehive increases the number of replicated objects per node as the Zipf parameter decreases to meet the targeted performance goal, and reduces the number of replicas as the Zipf parameter increases to reclaim space. The number of replicas per node shown in Figure 9 agrees with optimal solution provided by the analytical model. Overall, continuously monitoring and estimating the Zipf parameter of the query distribution enables Beehive to adjust the extent and level of replication to compensate for drastic, global changes.

Figure 8: Latency (hops) vs Time. This graph shows that Beehive quickly adapts to changes in the parameter of the query distribution and brings the average lookup performance to one hop.
\begin{figure}\centering\psfig{file=latency.alpha.eps, width=3in}\end{figure}

Figure 9: Objects stored per node vs Time. This graph shows that when the parameter of the query distribution changes, Beehive adjusts the number of replicated objects to maintain O(1) lookup performance with storage efficiency.
\begin{figure}\centering\psfig{file=objects.alpha.eps, width=3in}\end{figure}


next up previous
Next: Summary Up: Evaluation Previous: Flash Crowds
beehive-l@cs.cornell.edu 2004-02-11