memory from a remote server takes shorter than to the main memory cache from a
remote server as explained in Figure 5.4.
The increased cache hit ratio due to the NIC cache can be estimated using Equa
tion (5.3). In the exclusive scheme, the total number of cached data items (N
) in a
cluster is the summation of the cached data items in the main memory caches (N
and NIC caches (N
). While these numbers can vary by data item sizes, using the
average file size, we can estimate the total number of files in the main memory caches as
a, where a is the average file size and AS
is the accumulated
size of the main memory caches in a cluster. AS
can be calculated by N S
where N is the number of servers in the cluster and S
is the cache size in a server.
can be obtained by AS
a, where AS
is the accumulated size of
NIC caches in the cluster.
Figure 5.9 shows expected cache hit ratios of the original Web server and the
exclusive caching scheme, while the cache hit ratio is obtained using Equation (5.3).
Figure 5.9 indicates that the CSE trace benefits up to 8 nodes, and the UCB trace has
the best performance benefit from the exclusive caching scheme, while the PSU trace
has little performance gain. This is because the CSE and UCB traces have large data
set sizes (5.07GBytes and 4.6GBytes, respectively). However, even though the CSE
trace has a larger data set size than the UCB trace, the increased cache hit ratio by
the exclusive scheme in the UCB trace is twice higher than that of the CSE trace. The
reason is that the UCB trace has a lower popularity skewness ( ). In other words, the
requests from clients are distributed to broad data items compared to the other traces,
and thus incurring more disk accesses compared to the CSE trace with the same main