memory cache size. With the exclusive scheme, since we have an additional cache in the
NIC memory, many requests can be serviced by the NIC cache. Thus, it reduces the
number of disk accesses.
For the PSU trace, the performance benefit from the NIC cache is small. Note
that the PSU workload shows the highest popularity skewness, which implies that the
most frequently accessed files in the PSU trace may reside in the main memory cache
space of only one or two nodes. That explains why the increasing number of nodes does
not help to improve the cache hit ratio in both of the original and exclusive schemes.
On the other hand, the CSE and UCB traces show a continuing improvement of the
cache hit ratio when the number of nodes is increased up to 8 or 9 nodes compared to
the original cluster based Web server. Thus, we have two observations from this figure.
First, the performance improvement by the exclusive scheme is predictable. Second,
the number of server nodes in a cluster for obtaining a targeted cache hit ratio can be
Figure 5.10 shows throughput of the original and the exclusive NIC caching scheme
under the three examined traces. Figure 5.10(a) depicts that for the CSE trace the
exclusive scheme improves throughput up to 10% compared to the base model (no
caching case) because the trace contains many large data items and the NIC cache helps
in storing such files. For example, at node 3, the original server configuration has 9% of
the requests served by the disk, while using the exclusive caching, 4% of the requests are
served by the disk, and 7% of the requests are served by the NIC cache. the maximum
benefit from the exclusive scheme and this result matches with Figure 5.5. In addition,
Figure 5.10(b) indicates that the exclusive scheme shows more benefit as the number