Maximizing speedup through performance prediction for distributed shared memory systems

Abstract

Parallel applications executing on large-sized parallel systems achieve better speedup than on small-sized systems. However, because of the design and implementation of the distributed shared memory (DSM) system, there are some instances where large system size gives no further performance improvement over small system size. It is important to determine what system size will result in the maximum speedup while all kinds of applications are running on DSM systems. In this paper, we describe the design and implementation of the performance prediction mechanism in our DSM system, Proteus, which supports node reconfiguration to adjust the system size at run time. We adopt a simple computation model and combine it with run-time information to predict the performance under different system sizes. With this mechanism, it is possible to provide timely prediction results to adjust the size of the underlying system and thus maximize speedup.