Author : University of Michigan. Dept. of Electrical Engineering and Computer Science. Computer Science and Engineering Division
Publisher :
ISBN 13 :
Total Pages : 44 pages
Book Rating : 4.:/5 (36 download)
Book Synopsis Automatic Generation of Performance Bounds on the KSR1 by : University of Michigan. Dept. of Electrical Engineering and Computer Science. Computer Science and Engineering Division
Download or read book Automatic Generation of Performance Bounds on the KSR1 written by University of Michigan. Dept. of Electrical Engineering and Computer Science. Computer Science and Engineering Division and published by . This book was released on 1995 with total page 44 pages. Available in PDF, EPUB and Kindle. Book excerpt: Abstract: "Performance evaluation techniques have become critical indices for both architecture designers and software developers in recent years. Automatic tools are in demand for characterizing performance easily. In addition, a good performance analyzer should be able to point out the amount of performance loss relative to an ideal, where it occurs, and its causes. In this paper, MACS bounds, a hierarchy of bounds for modeling the performance, will be introduced for loop-dominated scientific code. This model successively considers the peak floating-point performance of a Machine of interest (M), the essential operations in a high level Application code of interest (MA), the additional operations in the Compiler-generated workload (MAC), the compiler-generated Schedule for this workload (MACS), and the actual delivered performance. In ascending through the MACS bounds hierarchy from the M bound, the model becomes increasingly constrained as it moves in several steps from potentially deliverable toward actually delivered performance. Each step from one bound to the next quantifies a performance gap associated with particular causes. The bound methodology will then be applied to a multiple-processor system to predict the run-time bound of a scientific application. We present three automatic bound generators, K-MA, KM̲ACSTAT and MACSG̲AP, based on our performance model on the KSR1 Massively Parallel Processing (MPP) machine. Finally, we will use the performance bound models and tools to experiment on a real parallel scientific application running on the KSR1. The analysis results of our selected loops will be illustrated by the MACS bounds hierarchy, machine utilization, and relative speed-up to uniprocessor. These information [sic] will be very useful for improving the performance of the application at the right point."