high performance computing on graphics processing units: hgpu.org

Evaluating the performance of parallel and heterogeneous programs and architectures can be challenging. An emulator or simulator can be used to aid the programmer. To provide guidance and feedback to the programmer, the simulator needs to present traces, reports, and debugging information in a coherent and unambiguous format. Although these outputs contain a lot of detailed information relative to the logical and physical transactions about the execution, they are usually extremely large and hard to analyze. What is needed is an interface into the simulator that can help programmers and architects shift through this myriad of data. In this contribution, we describe the M2S-Visual trace-driven visualization tool, a complementary addition to Multi2sim (M2S) heterogeneous system simulator. M2S-Visual provides a graphical representation of parallel program execution on the simulator. M2S is an established simulator, designed with an emphasis on running parallel applications on graphics processing units, and provides a number of instrumentation capabilities that enable research in architecture exploration and application characterization. This visualization framework, added to Multi2sim, aims to complement (and potentially replace) text-based statistical profiling, enabling the user to better understand each software transaction executed on the simulated hardware. While M2S supports emulation of both OpenCL and CUDA programs, our visualization framework presently only supports OpenCL execution. M2S supports execution on both CPUs (X86, ARM and MIPS) and GPUs (AMD Evergreen and Southern Islands, and NVIDIA Fermi and Kepler), but presently only supports detailed visualization on a multicore X86 CPU and AMD Evergreen and Southern Islands GPUs. Besides supporting OpenCL programming and debugging, an additional goal is to deliver a reliable product for teaching the details parallel programming execution on heterogeneous systems. Given the move to manycore architectures in the industry, this toolset is timely and addressing a growing gap in our educational infrastructure. The tool is also designed to support the research community, providing analysis of performance bottlenecks of OpenCL programs. We also introduce some new visualization which provide deeper insight into application performance and hardware resource utilization.