基于GPU的大规模多阶段任务系统可靠性并行计算方法

针对大规模多阶段任务系统(phased-mission system,PMS)的可靠性求解,引入并行计算思想,通过分析传统的一致化方法(uniformization method,UM),基于Nvidia提出的CUDA(compute unified device architecture)架构,实现了基于图形处理器(graphics processing unit,GPU)的UM并行算法(GPU-UM),并采用合并访问和共享内存技术,提高了GPU中数据负载的利用率;PMS中不同阶段参与任务的设备及其数量通常会发生变化,导致阶段间依赖性处理困难。通过对新设备加入、已有设备暂时退出任务或完全退出任务等3种基本情况的分析,提出了阶段间状态映射机制,实际中的阶段变化情况更加复杂,可综合上述3种基本情况进行处理。通过算例对比了GPU-UM、CUDA-UM、传统UM和Krylov子空间等4种算法的计算时间和可靠性结果,分析表明GPU-UM算法的计算耗时优于其他方法,且结果精度也能满足可靠性计算需求;同时,通过对比分析UM算法和Krylov子空间算法与仿真方法的结果误差,表明提出的阶段间映射机制能够正确处理PMS中阶段间的复杂依赖关系。

Reliability parallel computing method for large phased-mission system based on GPU

The parallel computing method is introduced for the mission reliability computation of large phased-mission system (PMS). According to analysis of the traditional uniformization method (UM), a UM parallel algorithm based on graphics processing unit (GPU) is given, which is denoted as GPU-UM. The GPU-UM algorithm is implemented under the compute unified device architecture (CUDA) put forward by Nvidia. The coherence visiting and shared memory techniques have been used to improve the data load utilization ratio on GPU. In addition, the component types and quantities involved in mission may be changed under different phases. The complex phase changes lead to difficulty about handling components dependency. A system states mapping mechanism is proposed by analyzing three basic situations: adding new components, some working components quitting transiently or completely. In practice, changes between two consecutive phases can be processed by coordinating the three situations above as it becomes more complicated. The computation time and reliability under the GPU-UM, the CUDA-UM, the traditional UM and the Krylov subspace algorithms are compared. Results show that the GPU-UM algorithm is the best among them both in computation time and accuracy. Furthermore, the computation errors between the two algorithms (UM and Krylov) and simulation method are analyzed. It shows that the proposed states mapping mechanism can handle the phases dependency in PMS efficiently.