43

3.6

Conclusions

In this chapter, we have investigated the impact of dynamic coscheduling algo 

rithms on cluster based Web server performance. We have modified a distributed Web

server model, called PRESS [23], to implement coscheduling techniques for remote cache

accesses to minimize the response time. We have developed a flexible and comprehensive

simulator testbed for evaluating different server architectures. The simulator is compre 

hensive in that it captures the characteristics of any user level communication at the

network level, models the behavior of a coscheduling algorithm at the kernel level, and

models the Web server details at the application level. The simulator is flexible enough

to analyze performance tradeoffs with various systems and workload parameters.

We compared four server designs (PRESS over VIA, PRESS with DCS, cosched 

uled PRESS with DCS and blocking, and Adaptive) on a 32 node cluster environment.

The simulation results indicate that it is possible to minimize the average response time

of client requests by coscheduling the communicating processes at remote nodes. In par 

ticular, the use of the DCS coscheduling scheme reduced the latency up to four times in

lightly loaded servers compared to that with the original PRESS model. However, the

performance of the dcs model degrades as the probability of coscheduling communicat 

ing processes decreases due to competition among the remote cache read requests. Thus

we proposed the adaptive model. The adaptive model can provide a stable performance

benefit by employing a coscheduling algorithm selectively depending on the systems load.