Characterization of ILP Distribution for NASA NAS Parallel Benchmarks

A characterization study of analyzing dynamic instruction traces to characterize program parallelism is conducted. This study supports that the experimental design of supercomputer and parallel computers calls for quantifiable methods to evaluate the requirements of different workloads within an application domain. Such methods can help establish the basis for scientific design of parallel computers driven by application needs, to optimize performance to cost. In addition, the application characteristics can be used early in the design process to identify bottlenecks such as not having enough resources, not having enough parallelism in the instruction stream, or using a too restrictive scope of concurrency detection. The selection of which features to include in a new computer system depends on the needs of the workloads the system will execute. The number and type of functional units are among the most important design decisions. Therefore, this paper presents an instruction-level characterization for analyzing dynamic traces using a trace-driven simulator. It investigates the parallel system needs for a class of contemporary benchmarks taken from NASA/NAS Parallel Benchmarks (NPB) suite. The NPB represents an implementation independent problem set, representative of Computational Aeroscience workload computations. The NPB workloads have been implemented on nearly every parallel platform and results have been reported by the vendors. Data is presented for NBP requirements of these resources. The requirements suggest upper limits on the resources needed for efficient processors. In this study, we also examine non-uniformities in the distribution of instruction-level parallelism. Several nonuniformities in instruction-level parallelism are investigated including variation between benchmark class and by instruction class within benchmark. In addition, the average instruction class distribution as well as the shortest path a workload would be executed on a parallel machine will be shown. The results confirm that workloads in NPB represent a wide range of non-redundant applications with different characteristics.