41

Figure 3.8 (c) and (d) show the latency results when the servers are heavily 

loaded. When the cluster is heavily loaded, the disk accesses easily become a performance

bottleneck. Since the required time for a disk access is much longer than the local or

remote cache access, the average latency time jumps when the number of disk accesses

increases. Figure 3.8 (c) shows that the average response time of the three models soars at

the point R = 40%. Figure 3.8 (d) shows that these disk accesses become an increasingly

severe performance bottleneck when the Web servers receive more requests. Until R is

30%, the dcs and adaptive models provide stable performance. After this point, the

average latency increases fast. The dcs model shows the worst performance in Figure 3.8

(d) from R = 5% to R = 30% due to steady increase of remote cache read requests. This

result agrees with the previous experiment, which showed that the performance of dcs is

degraded when the load increases. By increasing the replication ratio, we can gain the

performance benefit due to the large number of local cache hits. The proper replication

percentage varies with the system load. However, the disk becomes a severe performance

bottleneck when a system is loaded heavily. Therefore, reducing the number of disk

accesses is more important than increasing the number of local cache hits to provide

better performance in a cluster.