Next, we examine the performance of proactive and passive caching in response to changes in object popularity. We modify the trace to suddenly reverse the popularities of all the objects in the system. That is, the least popular object becomes the most popular object, the second least popular object becomes the second most popular object, and so on. This represents a worst case scenario for proactive replication, as objects that are least replicated suddenly need to be replicated widely, and vice versa. The switch occurs at t=40, and we issue queries from the reversed popularity distribution for another 40 hours.
Figure 6 shows the lookup performance of Pastry, PC-Pastry and Beehive in response to flash crowds. Popularity reversal causes a temporary increase in average latency for both Beehive and PC-Pastry. Beehive adjusts the replication levels of its objects appropriately and reduces the average lookup performance to about 1 hop after two replication intervals. The lookup performance of passive caching also decreases to about 1.6 hops. Figure 7 shows the instantaneous rate of object transfer induced by the popularity reversal for Beehive and PC-Pastry. The popularity reversal causes a temporary increase in the object transfer activity of Beehive as it adjusts the replication levels of the objects appropriately. Even though Beehive incurs this high rate of activity in response to a worst-case scenario, it consumes less bandwidth and imposes less aggregate load compared to passive caching.
 |
 |