TARUC: A Topology-Aware Resource Usability and Contention Benchmark.

Computer architects have increased hardware parallelism and power efficiency by integrating massively parallel hardware accelerators (coprocessors) into compute systems. Many modern HPC clusters now consist of multi-CPU nodes along with additional hardware accelerators in the form of graphics processing units (GPUs). Each CPU and GPU is integrated with system memory via communication links (QPI and PCIe) and multi-channel memory controllers. The increasing density of these heterogeneous computing systems has resulted in complex performance phenomena including non-uniform memory access (NUMA) and resource contention that make application performance hard to predict and tune. This paper presents the Topology Aware Resource Usability and Contention (TARUC) benchmark. TARUC is a modular, open-source, and highly configurable benchmark useful for profiling dense heterogeneous systems to provide insight for developers who wish to tune application codes for specific systems. Analysis of TARUC performance profiles from a multi-CPU, multi-GPU system is also presented.