An Exploration of Performance Attributes for Symbolic Modeling of Emerging Processing Devices

Abstract

Vector, emerging (homogenous and heterogeneous) multi-core and a number of accelerator processing devices potentially offer an order of magnitude speedup for scientific applications that are capable of exploiting their SIMD execution units over microprocessor execution times. Nevertheless, identifying, mapping and achieving high performance for a diverse set of scientific algorithms is a challenging task, let alone the performance predictions and projections on these devices. The conventional performance modeling strategies are unable to capture the performance characteristics of complex processing systems and, therefore, fail to predict achievable runtime performance. Moreover, most efforts involved in developing a performance modeling strategy and subsequently a framework for unique and emerging processing devices is prohibitively expensive. In this study, we explore a minimum set of attributes that are necessary to capture the performance characteristics of scientific calculations on the Cray X1E multi-streaming, vector processor. We include a set of specialized performance attributes of the X1E system including the degrees of multi-streaming and vectorization within our symbolic modeling framework called Modeling Assertions (MA). Using our scheme, the performance prediction error rates for a scientific calculation are reduced from over 200% to less than 25%.