Are all scientific workloads equal?

Abstract

Widely-used benchmarks are commonly classified as either scientific or commercial. Although process execution characteristics have been used as indicators of a benchmark's classification, a set of these characteristics along with a mechanism that can be used to easily compare and contrast workloads and partition them into classes with respect to these characteristics has not been identified. This paper identifies a set of process execution characteristics (PEC) that can be used to compare and contrast workloads and a method that can be used to partition workloads with respect to their PEC. These PEC, such as instruction locality, execution cycles per instruction, and context-switch frequency, are displayed with a high-density visualization tool called the PEC-Graph. Using the centroid linkage algorithm, processes' PEC are partitioned into clusters that are used to construct a taxonomy of workloads that is finer grained than taxonomies previously reported in the literature. The finer-grained categorization of workloads enables computer architects to select workloads that are known to stress specific architectural features, yielding potentially better performance analysis of new designs.