Performance Considerations : A Case Study Using a Scalable Shared-Virtual-Memory Machines

Abstract

Recent trends in parallel processing suggest that the issue of performance prediction is becoming more complex and dif- ficult. Scientists have adopted massively parallel computing as a cost-effective way to achieve high computing power. They have introduced various architectures and algorithms to deliver performance scalability with many processors. Shared virtual memory and other kinds of system support, which hide the communica- tion and other implementation details from the user, are becoming more prevalent. At the same time, with various architectures and algorithms available, performance prediction is becoming critical in choosing an appropriate algorithm-machine pair for an application, especially when the machine has a sophisticated hierarchical architecture. In this article, we combine simple formulas with runtime information to predict performance in modern parallel computers. After presenting a simple prediction formula, we discuss a case study involving a virtual memory machine to illustrate how to use the formula in practice. We dis- cuss four different aspects: • We propose a method to measure the needed runtime parameters. • We propose an adjustment to catch the influence of architecture vari- ation when the system size is scaled up from one level of architecture hierarchy to another. • We demonstrate, through the case study, that it is possible to predict the influence of architecture hierarchy on scalability by simply using hardware specifications. Finally, we discuss the issue of choosing an appropriate algorithm for a given application when the computing system is scaled up from one level of hierarchy to another.

Links