Research On Parallel Computing Model For CPU/GPU Heterogeneous System

The rapid development of the national economy and technology put forward higher requirement on the performance of HPC(High-performance computer). The traditional method to develop HPC through pure CPU faces tremendous challenge in energy consumption, heat dissipation and cost. Heterogeneous architecture combines the advantage of general processor and accelerator, and gradually becomes the mainstream architecture in HPC domain. GPU has been widely used in developing heterogeneous computer system for its great advantages in performance, storage bandwidth, power consumption and programmable properties. In the era of green computing, the CPU/GPU heterogeneous HPC systems have a very promising future. Large CPU/GPU heterogeneous systems have immense computing power, which provides a good opportunity for large-scale scientific and engineering applications. However, the complex hardware structure and specific execution scheme present a huge problem for parallel computing researchers.As a branch of the parallel computing, parallel computing model is an abstraction of the underlying parallel computer systems, reflects their resources and performance characteristics with several parameters. It builds a bridge between software and hardware and it’s of great value to parallel algorithm designers. Generally speaking, The parallel computing model is generally suitable for many platforms, However, the CPU/GPU heterogeneous HPC systems, which include three levels of heterogeneous characteristics and have specific structural characteristics and performance factors that cannot be accurately described by the current models. Hence, it is urgent to carry out parallel computing research on such HPC platform. It can provide support for the current and the future parallel application development based on these platforms. Therefore, studies on parallel computing model of heterogeneous CPU/GPU clusters is of significant importance.The major work in this article is listed as follows:1) We perform a systematic survey of parallel computing model and discuss the characteristics and the future of parallel computer architectures, then we summarize the research progress at home and abroad about parallel computing model. Based on the current work, various parallel computing models are compared and analyzed.2) We propose a parallel computing model for large CPU/GPU heterogeneous high-performance computing systems, which was used to abstract these systems’computing power and communication behaviors. Based on the analysis of the new model’s complexity, we map the model to the TH-1A system, and all the model’parameters in the TH-1A platform are determined. 3) We use CUDA language to implement IS and FT in the NAS Parallel Benchmark, and effectively apply them to the TH-1A platforms. Through benchmarks such as communication-intensive application NPB-IS and compute-intensive application NPB-FT, we use real-time overhead to verify the proposed new model.4) We optimize some of typical scientific applications after using gprof tool to identify hotspots of NPB-IS and NPB-FT, then optimize the computing and communication patterns of the NPB-IS, NPB-FT through CPU/GPU heterogeneous cluster programming optimization method, and a detailed comparison and analysis of the final results are listed.