17

validated against those from measurements on the 16 node cluster. This provided us

with an accurate platform to experiment with server design alternatives.

We analyze four server designs (PRESS over VIA, coscheduled PRESS with DCS,

coscheduled PRESS with DCS and blocking, and Adaptive) using a 32 node configura 

tion. Our results shows that the DCS coscheduling algorithm can reduce the average

latency up to four times in lightly loaded servers compared to the PRESS over VIA

model. The adaptive model yields a gain in performance by applying the DCS coschedul 

ing algorithm selectively depending the server load. With the DCS Blocking scheme, the

latency reduction is less because of the blocking overhead.

The rest of this chapter is organized as follows. In Section 2, we provide a sum 

mary of the PRESS model and coscheduling techniques. Section 3 and Section 4 outline

the proposed coscheduled server to the PRESS Web model and the simulation plat 

form respectively. The performance results are analyzed in Section 5, followed by the

concluding remarks in the last Section.