Comparing the Performance of the DASH and Cedar Multiprocessors for Scientific Applications

this paper, we compare the performance of DASH and Cedar under a set of varied parallel scientific loads. We focus on three main differences between the two machines: cache coherence supported in hardware (DASH) versus managed by the programmer (Cedar), task-based parallelism (DASH) versus loop-based parallelism (Cedar), and the hardware support to exploit cooperation among the processors in a cluster. In our analysis, we compare performance differences to the extra hardware or programming cost involved. Our results indicate that, for several common mediumgrain scientific applications, lack of hardware support for cache coherence is not a major problem. These codes are regular enough for the programmer to easily manage memory. We also note that machine programmability is enhanced if the machine supports both loop-level parallelism and pairwise processor synchronization via locks. Finally, support for processor clusters boosts the performance by 70-80% at a moderate cost. The paper is organized as follows. Section 2 discusses the differences between DASH and Cedar. Section 3 presents the applications. Sections 4-7 compare the machines using microkernels, parallelized serial codes, parallel codes ported from one machine to the other, and an application that exploits clusters