Accelerating HPC Applications Using Computational Storage Devices

During the past decades, computers have become more powerful, and the size of the problems they can solve have increased drastically. Today's supercomputers have millions of processors orchestrated together to create an environment for solving superscale problems. On the other hand, the power consumption of modern supercomputers has increased to thousands of kWs and data movements account for a large portion of this increment. In this paper, we show how computational storage technology -running applications inside storage devices-provides a scalable solution for running HPC applications and improves performance while decreasing energy consumption. We introduce Catalina, a high-capacity Computational Storage Device (CSD) that is capable of running a wide range of applications in-place by utilizing a quad-core 64-bit ARM processor coupled with NEON SIMD engines. As the proof of concept, we built a fully-functional Catalina prototype and a server including six Catalina CSDs to run HPC applications. The results show up to 2.2x improvement in performance and 54% reduction in energy consumption for running multi-dimensional FFT benchmarks on different datasets.