73

To study the performance impact of the proposed NIC caching schemes on Web

server cluster based on SAN, we conduct a performance analysis based on Web work 

loads with different characteristics. We first obtain three different Web server logs (i.e.,

Penn State CSE (CSE) [63], UC Berkeley (UCB) [71] and Penn State University (PSU)

[62] traces) for analyzing of the characteristics of Web content. These traces were gath 

ered recently (except for the UCB trace) and contain a large amount of data. Through

our analysis, we found that these three traces show some major differences in statis 

tical characteristics and accessed patterns. These differences meet our goal of explor 

ing performance implications of the proposed caching schemes under different workload

characteristics. As such, the findings drawn from this study can be used for designing

alternative caching schemes and for tuning the caching schemes to maximize Web server

performance.

We conduct an extensive performance evaluation on a simulation based Web server

(with or without NIC caching). Our experimental results show that the NIC caching

schemes significantly increase throughput over the original system. The exclusive caching

scheme becomes beneficial when workload shows a low popularity skewness. The inclu 

sive caching scheme is particularly effective with workloads with large file sizes and high

popularity skewness. We also examine the performance impact of the I/O intercon 

nection architectures and of the scalability of cluster size. Our results show that the

NIC exclusive caching scheme improves system performance even if the bandwidth made

available by intra cluster communication technologies continues to increase. Also, NIC

caching schemes provide scalable throughput with various cluster sizes.