System-level modelling for performance estimation of reconfigurable coprocessors

Abstract

The capabilities of general-purpose workstations are commonly enhanced by the addition of application-specific coprocessors located on the system expansion bus or a dedicated local bus. In order to determine the limits of applicability of such systems, performance estimation tools are required which are capable both of generating accurate predictions and supporting rapid evaluation of architectural alternatives. This paper describes a performance estimation method which meets these requirements. By combining data acquired using a variety of established techniques, and integrating these using a novel approach designed to capture far more of the intrinsic complexity of the system, extremely accurate estimates of performance can be generated. A detailed uncertainty analysis is provided, and the method is evaluated by comparing an application running on a coprocessor-enhanced workstation with model-based predictions, demonstrating estimation accuracy to be within □5% of measured values. © Springer-Verlag Berlin Heidelberg 2002.