A Multi-Tiered Optimization Framework for Heterogeneous Computing

Modern computing nodes often contain more than just a CPU. With the advent of GPU accelerators and Xeon Phi co-processors, there are many architectures available for data processing. However, it is difficult to understand which device is best for a given application. The issue of real-world performance originates in the lack of quantifiable data and method for analysis. This paper presents a novel, multi-tiered framework that leverages Pareto optimization to objectively construct the best processing node for a set of computational kernels. By deconstructing the optimization process into three distinct framework tiers (kernel, device, and system), the system designer is able to understand how the various computational variables impact device choices. We show how we leverage a combination of metrics and benchmarking to form various Pareto sets. Moving through the tiers, these Pareto sets are combined to identify the various combinations that enable maximum performance.