reducing the number of DMA transfers and PCI overhead, the benefit of the NIC caching
scheme increases with high intra cluster communication.
Impact of Cache Replacement Policy
In the inclusive NIC caching scheme, we expect good performance when the NIC
cache contains the most frequently accessed files. However, implementation of perfect
LFU is not trivial. Thus, in our experiment, we use the LRU cache replacement policy
which is most widely used, to maintain the N most recently used files in the NIC cache.
Copying the most recently accessed files from the main memory cache to the NIC cache
incurs expensive I/O cost. Thus, we do not update the NIC cache as frequent as the
main memory cache is updated. Instead, the NIC cache is updated periodically (i.e., the
default is every 20 second).
Here, we investigate the impact of the NIC cache update interval on performance.
The result with the UCB trace in Figure 5.15 shows that the throughput is almost un
affected as the update interval is varied from 10 to 200 seconds (the result of the CSE
trace is omitted since it shows the same result with the UCB trace). This is because the
access pattern of Web data items follows the Zipf distribution  (i.e., most requests are
concentrated to a small portion of the data set and the 256 MBytes NIC cache is large
enough to hold most of highly accessed files.). In addition, the static Web content is
rarely updated. Thus, the NIC cache content is almost unchanged between two updates,
when we use the inclusive NIC scheme to cache the most recently accessed files on the
NIC memory. Thus, based on our experiments, we conclude that the update interval