The Design Of GPU Cluster Oriented To General Purpose Computing

Because of powerful computing power, high-speed memory access bandwidth and supporting large-scale data-level parallelism, the GPUs have become the mainstream accelerators insupercomputers and high performance computing field.More and more applications achieve substantial speedup via GPU-based data-level parallelism re-designre. Therefore, in the HPC field, GPU clusters have become a new important research hot-spot, and the traditional clusters have being evolving to become a GPU-enhanced clusters or GPU clusters.From the view of architecture, the architecture feature of GPU clusters is summarized through the analysis of traditional clusters and GPU features. The paper points out heterogeneous computing characteristic and the powerful capability of multi-level parallel computing in GPU clusters. A heterogeneous GPU cluster is built by the integration of existing computing resources under the guidance of theoretical studies.This paper describe the build step of the GPU cluster in detail.Mainstream programming models for GPU clusters is a simple combination of message passing mechanism and heterogeneous computing. This approach is often inefficient and error-prone. The root cause of the pitfall in the description and design of the parallel application in these model is that the programming and computational models can not fit the many-core architecture of GPU and the cluster environment. Stream programming model analyse in depth the parallel computing characteristics from the top programming model and to provide an explicit multi-level parallelism design, and it can guide the programmers’perspective from the simple division of functions and processes to the division of data-related computing. Therefore, Stream programming model can fully fit the architecture of GPU clusters and exert the powerful data-parallel computing power for large-scale data.GPU-enhanced clusters as mainstream components in the HPC field are expected to be heterogeneous in system layer and node layer as the evolvement of processing elements (CPUs and GPUs) and the expansion of nodes. In this paper, we proposed an energy efficient task scheduling scheme for heterogeneous tasks in the heterogeneous GPU-enhanced clusters. A system model, a task model and an energy evaluation model for the heterogeneous clusters are formulated in this paper. According to the particular node selection policy, it can decrease the static energy consumption of GPUs in idle status. By the division of task types and buddy allocation, it can improve the utilization of the CPU resource to increase energy efficiency. It can change the dynamic voltage and frequency scaling feature in accordance to the load of system. What’s more, the scheme is friendly and compatible with algorithm-level and instruction-level energy optimization for GPUs and can be used along with them to achieve better energy efficiency.The paper studies deeply the architecture, programming model, the energy efficiency of GPU clusters and provides a set of design for the GPUcluster, and guide the completion of the structures of the hardware and software environment in the built GPU cluster. We implement and deploy the parallel programming environment which is suitable for stream processing applications based on large-scale data-level parallelism of GPU clusters. Finally, an ernegy efficient task scheduling scheme based on node heterogeneous GPU clusters is proposed from the system level. The paper can guide the design of the GPU clusters which is oriented to general purpose computing.