Parallelism and Data Movement Characterization of Contemporary Application Classes

摘要

This paper presents a framework for characterizing the distribution of fine-grained parallelism, data movement, and communication-minimizing code partitions. Understanding the spectrum of parallelism available in applications, and how much data movement might result if such parallelism is exploited, is essential in the hardware design process because these properties will be the limiters to performance scaling of future computing systems. The framework is applied to characterizing 26 applications and kernels, classified according to their dominant components in the Berkeley dwarf/ computational motif classification. The distributions of ILP and TLP over execution time are studied, and it is shown that, though mean ILP is high, available ILP is significantly smaller for most of the execution. The results from this framework are complemented by hardware performance counter data on two RISC platforms (IBM Power7 and Freescale P2020) and one CISC platform (Intel Atom D510), spanning a broad range of real machine characteristics. Employing a combination of these new techniques, and building upon previous proposals, it is demonstrated that the similarity in available ideal-case parallelism and data movement within and across the dwarf classes, is limited.