high performance computing on graphics processing units: hgpu.org

The High Performance Computing (HPC) community recognizes energy consumption as a major problem. Extensive research is underway to identify means to increase energy efficiency of HPC systems including consideration of alternative building blocks for future systems. This thesis considers one such system, the Texas Instruments Keystone II, a heterogeneous Low-Power System-on-Chip (LPSoC) processor that combines a quad core ARM CPU with an octa-core Digital Signal Processor (DSP). It was first released in 2012. Four issues are considered: i) maximizing the Keystone II ARM CPU performance; ii) implementation and extension of the OpenMP programming model for the Keystone II; iii) simultaneous use of ARM and DSP cores across multiple Keystone SoCs; and iv) an energy model for applications running on LPSoCs like the Keystone II and heterogeneous systems in general. Maximizing the performance of the ARM CPU on the Keystone II system is fundamental to adoption of this system by the HPC community and, of the ARM architecture more broadly. Key to achieving good performance is exploitation of the ARM vector instructions. This thesis presents the first detailed comparison of the use of ARM compiler intrinsic functions with automatic compiler vectorization across four generations of ARM processors. Comparisons are also made with x86 based platforms and the use of equivalent Intel vector instructions. Implementation of the OpenMP programming model on the Keystone II system presents both challenges and opportunities. Challenges in that the OpenMP model was originally developed for a homogeneous programming environment with a common instruction set architecture, and in 2012 work had only just begun to consider how OpenMP might work with accelerators. Opportunities in that shared memory is accessible to all processing elements on the LPSoC, offering performance advantages over what typically exists with attached accelerators. This thesis presents an analysis of a prototype version of OpenMP implemented as a bare-metal runtime on the DSP of a Keystone I system. An implementation for the Keystone II that maps OpenMP 4.0 accelerator directives to OpenCL runtime library operations is presented and evaluated. Exploitation of some of the underlying hardware features of the Keystone II is also discussed. Simultaneous use of the ARM and DSP cores across multiple Keystone II boards is fundamental to the creation of commercially viable HPC offerings based on Keystone technology. The nCore BrownDwarf and HPE Moonshot systems represent two such systems. This thesis presents a proof-of-concept implementation of matrix multiplication (GEMM) for the BrownDwarf system. The BrownDwarf utilizes both Keystone II and Keystone I SoCs through a point-to-point interconnect called Hyperlink. Details of how a novel message passing communication framework across Hyperlink was implemented to support this complex environment are provided. An energy model that can be used to predict energy usage as a function of what fraction of a particular computation is performed on each of the available compute devices offers the opportunity for making runtime decisions on how best to minimize energy usage. This thesis presents a basic energy usage model that considers rates of executions on each device and their active and idle power usages. Using this model, it is shown that only under certain conditions does there exist an energy-optimal work partition that uses multiple compute devices. To validate the model a high resolution energy measurement environment is developed and used to gather energy measurements for a matrix multiplication benchmark running on a variety of systems. Results presented support the model. Drawing on the four issues noted above and other developments that have occurred since the Keystone II system was first announced, the thesis concludes by making comments regarding the future of LPSoCs as building blocks for HPC systems.