40

N

H

rep C

Local

=

i=1 i 

+ H

remote

1

n

H

disk

= 1 - (H

remote

+ H

local

).

Therefore, as the replication ratio increases, H

Local

increases, but H

remote

de 

creases because the total number of the cached files in the cluster decreases causing many

disk accesses. If the replication ratio becomes small, many requests should be served as

remote cache read, and thus it reduces the number of disk accesses.

Figure 3.8 shows the average latency as a function of the replication percentage

(R), when the cluster is loaded differently. Figure 3.8 (a), (b), (c) and (d) are the

latency results when k is set to 10, 8, 6 and 4 respectively, where k is the parameter

of the Pareto distribution for the incoming request interval. In Figure 3.8, the Y axis

represents the average latency and the X axis represents the replication ratio in the

cluster. Figure 3.8 (a) and (b) show the latency results when the servers are lightly 

loaded. These figures show that as the replication percentage increases, the latency of the

three models decreases slightly, and then increases. As shown above, as the percentage of

replication increases, the local cache hit ratio increases, but the accumulated cache size

decreases resulting in additional disk accesses. The press via model benefits more than

the adaptive and dcs models do because in the press via model, the remote cache read

latency is much longer than the local cache read latency. Thus, the increase of local

cache hits helps to reduce the average latency time in the press model. However, the

latency of the three models increases slightly as the replication ratio increases due to

the additional disk accesses. But when the cluster is lightly loaded, the performance

degradation due to the disk access is not very significant.