16

implementation based, are limited in terms of scalability and other system parameters.

In addition, implementation of coscheduling is a substantial effort, and it may be worth 

while to estimate the achievable benefits prior to undertaking an implementation based

study.

In this thesis, we propose a coscheduled PRESS model that can employ a coschedul 

ing algorithm to facilitate faster communication. The remote cache send and the re 

mote cache recv threads are used to handle requests to a remote cache. To minimize

the response time for the requests, which need the intra cluster communication, the

coscheduling scheme is applied. The rationale for this is that since many requests from

clients should be served as remote cache read in the PRESS Web server, it is critical to

minimize the response time of these requests.

We have developed a flexible and accurate simulation testbed for analyzing differ 

ent server architectures. The simulator, written in CSIM, has three major components.

First, it captures the network/SAN behavior along with the VIA [30] functionalities.

Second, it can model any of the prior coscheduling algorithms and estimate the cor 

responding delay. Finally, it models a distributed server architecture and computes

the overall throughput and response time for various system configurations. In this re 

search, we have evaluated two coscheduling alternatives, dynamic coscheduling (DCS)

[69] and DCS with immediate blocking. We use the network and scheduling timing pa 

rameter values in the simulator (one way latency, processing overheads in NIC, context

switching overhead and interrupt overhead, etc.) from an actual implementation of var 

ious coscheduling algorithms in a 16 node Linux cluster connected via a Myrinet switch

[27, 3]. The round trip latencies from the simulator, under different settings, were first