Subjective versus objective: classifying analytical models for productive heterogeneous performance prediction

Heterogeneous analytical models are valuable tools that facilitate optimal application tuning via runtime prediction; however, they require several man-hours of effort to understand and employ for meaningful performance prediction. Consequently, developers face the challenge of selecting adequate performance models that best fit their design goals and level of system knowledge. In this research, we present a classification that enables users to select a set of easy-to-use and reliable analytical models for quality performance prediction. These models, which target the general-purpose graphical processing unit (GPGPU)-based systems, are categorized into two primary analytical classes: subjective-analytical and objective-analytical. The subjective-analytical models predict the computation and communication components of an application by describing the system using minimum qualitative relations among the system parameters; whereas the objective-analytical models predict these components by measuring pertinent hardware events using micro-benchmarks. We categorize, enhance, and characterize the existing analytical models for GPGPU computations, network-level, and inter-connect communications to facilitate fast and reliable application performance prediction. We also explore a suitable combination of the aforementioned analytical classes, the hybrid approach, for high-quality performance prediction and report prediction accuracy up to 95 % for several tested GPGPU cluster configurations. The research aims to ultimately provide a collection of easy-to-select analytical models that promote straightforward and accurate performance prediction prior to large-scale implementation.