Characterization of GPGPU Workloads via Correlation-Driven Kernel Similarity Analysis

Graphics Processing Units are emerging as a general-purpose high-performance computing devices (GPGPUs). Although this has led the creation of numerous GPGPU workloads available, there is a lack of a systematic approach to characterize GPGPU-applications. This paper proposes a similarity-based methodology for the characterization of GPU workloads. The proposed methodology successfully characterizes GPGPU workloads using kernel signatures and clustering algorithms. The signatures are derived from architecture-aware features of the workload, in particular from hardware performance counters. The evaluation of the proposed characterization approach includes a diversity of GPU benchmark suites such as Nvidia CUDA SDK, Parboil and Rodinia.