cactus : top-down GPU-compute benchmarking using real-life applications

Benchmarking is the de facto standard for evaluating hardware architectures in academia and industry. While several benchmark suites targeting different application domains have been developed for CPU processors over many decades, benchmarking GPU architectures is not as mature. Since the introduction of GPUs for general-purpose computing, the purpose has been to accelerate (a) specific part(s) of the code, called (a) kernel(s). The initial GPU-compute benchmark suites, which are still widely used today, hence consist of relatively simple workloads that are composed of one or few kernels with specific unambiguous execution characteristics. In contrast, we find that modern-day real-life GPU-compute applications are much more complex consisting of many more kernels with differing characteristics. A fundamental question can hence be raised: are current benchmark suites still representative for modern real-life applications? In this paper, we introduce Cactus, a collection of widely used real-life open-source GPU-compute applications. The aim of this work is to offer a new perspective on GPU-compute benchmarking: while existing benchmark suites are designed in a bottom-up fashion (i.e., starting from kernels that are likely to perform well on GPUs), we perform GPU-compute benchmarking in a top-down fashion, starting from complex real-life applications that are composed of multiple kernels. We characterize the Cactus benchmarks by quantifying their kernel execution time distribution, by analyzing the workloads using the roofline model, by performing a performance metrics correlation analysis, and by classifying their constituent kernels through multi-dimensional data analysis. The overall conclusion is that the Cactus workloads execute many more kernels, include more diverse and more complex execution behavior, and cover a broader range of the workload space compared to the prevalently used benchmark suites. We hence believe that Cactus is a useful complement to the existing GPU-compute benchmarking toolbox