Research On GPU Parallel Computing And Application For HPC Cloud

In recent years, GPU(Graphics Processing Unit) parallel computing technology has become a hot topic in the field of high performance computing(HPC). The great floating-point computational power of GPUs provides good support for large-scale scientific and engineering computing. In addition to the traditional HPC applications, the emerging high-performance computing applications are also increasingly demanding for high performance. In the aspect of user-oriented services, the traditional HPC faces many problems: how to provide users with a flexible service model that allows users to self-manage computing resources; how to provide users dynamical and scalable computing resources. As a type of HPC resource management and service model based on cloud computing, HPC cloud solves the problem in the aspect of customer service for the traditional HPC.Currently, both HPC cloud and GPU parallel computing are rather popular research fields in computer application technology. In the intersecting part of these two hotspots, the study of GPU parallel computing technology and related applications in the HPC cloud environment are still at starting stage. For HPC cloud computing scheduling the GPU computing resources and GPU communication performance degradation are great challenges to the study of GPU parallel computing technology. Hence, it’s necessary to focus on HPC cloud and carry out research on GPU computing technology, so as to provide support for building HPC cloud platform based on GPU and developing of GPU parallel application on this platform in the future.For the GPU parallel computing technology in the environment of HPC cloud, this thesis discusses the scheduling of computing resources and the communication among multi-GPU in the environment of HPC cloud. Thus, the study of related GPU parallel computing technology application is carried out. The research work and results of this thesis are mainly about:1) A multi-GPU computing resource scheduling mechanism is proposed for HPC cloud. Resource scheduling is one of the key issues in the Cloud Computing. As a special computing resource in HPC Cloud, GPU’s scheduling has its particularity. In the proposed scheduling mechanism, delaying between nodes and internal data transmission in node of computing tasks are taken into account. In order to take advantage of GPU computing resources and hide the transmission delay, a model of GPU computing resources is built. The model describes the three-stage scheduling processes of "Transmission& Transmission& Execution". To reduce tasking pressure on the central node, a tree structure based GPU computing resources retrieval is proposed for HPC Cloud environment. Experimental results show that the proposed scheduling mechanism has many advantages in improving resource utilization and quality of service.2) A low delay GPU communication mechanism for HPC cloud is proposed. The GPU-equipped HPC Cloud meets the requirements of multiuser online data processing, while reducing the performance drawback due to virtualization and achieving efficient communication between GPU computing resources. Using data-hold method, the purpose of GPU communication strategies for HPC cloud is to reduce the communication delay during the calculation with GPU in a virtualized environment, and reduce the performance loss coming with virtualization and provide users with highly scalable computing services. In addition, based on a type of data streaming transport, data high reuse strategy for real-time data is purposed further. The analysis for low delay communication and data reuse strategies are from two aspects: communication effectiveness and scalability of platform.3) As an application of I/O bound in HPC cloud environment, this thesis studies quantum search algorithm simulation using GPU with feature of single-instruction multi-threaded and proposes a simulation method for quantum search algorithm in GPU HPC cloud platform. One of the proposed methods focuses on the data compressing and improves the GPU memory utilization, which enlarge the simulation rang. The other proposed method focuses on the general simulation of quantum search algorithm, which solves the problem of calculating and extracting intermediate variables. The advantages and disadvantages of proposed simulation methods are compared and analyzed through experiments.4) As an application of compute-bound in HPC cloud environment, this thesis studies a three dimensional atomic resolution tomography reconstruction method, which is implemented on GPU HPC cloud platform efficiently. A cylindrical Fourier transform based three-dimensional reconstruction algorithm is proposed. Then, in order to achieve high accuracy for nonequispaced Fourier transform, GPU parallel computing method is proposed. In the method, the convolution step is optimized on GPU. In order to achieve high efficiency in on-chip memory usage, pre-computed compressed datasets recording the relationship of thread and output data is applied through input-driven approach, which helps to avoid the write-conflict. The performance was measured by comparing with the available GPU version. The test demonstrates an improved performance in single precision and double precision, respectively. Furthermore, to illustrate the availability of the reconstruction method in HPC cloud with GPU platform, the test evaluate the method performance for reconstructing a three dimensional gold nanoparticle using the experimental data from University of Illinois at Urbana-Champaign.