This policy is scalable unlike the single front end layer 7 webswitch. The PRESS Web
server is the first effort to analyze the impact of user level communication on server per
formance. It has been shown that user level communication mechanisms such as Virtual
Interface Architecture (VIA)  can significantly minimize the communication over
head. The recently proposed InfiniBand Architecture (IBA) , for designing SANs for
clusters, has adopted the VIA design paradigm. Several commercial implementations of
VIA have demonstrated the advantage of user level communication [19, 54, 38, 51, 39].
These previous studies mainly focused on system throughput. However, in addi
tion to high throughput, low latency is another objective function for designing cluster
based Web servers . The steady growth of dynamic Web content also makes latency
reduction more critical. In addition to user level communication, dynamic coschedul
ing of communicating processes has been used as an effective approach to minimize the
overall execution time of parallel applications on clusters [14, 53, 69]. As the PRESS
model and other distributed server models indicate significant amount of intra cluster
communication to service client requests and update caching information, coschedul
ing of communicating processes between the nodes may help in minimizing the latency.
The motivation of this work is to examine this philosophy by utilizing both user level
communication and coscheduling on cluster based servers.
Unlike some other studies, however, we resort to simulation instead of an actual
implementation. This is primarily because a detailed simulator is better suited to evalu
ate the impact of various parameters such as the cluster size, network type, communica
tion model, coscheduling model, node configuration in terms of cache size and processor
speed, and various workload parameters. On the other hand, prior studies which are