GPU机群系统上加速EMAN(英文)  

Accelerating EMAN on A GPU Cluster System

作　　者：李兴建[1] 谭光明[1] 郭子昱[1] 

机构地区：[1]中国科学院计算技术研究所,北京100084

出　　处：《生物物理学报》2010年第7期600-605,共6页Acta Biophysica Sinica

基　　金：supported by grants from National Natural Science Foundation of China (60803030, 60633040);Chinese Academy of Sciences (KGCX1-YW-13)~~

摘　　要：近年来,高质量的三维重构对计算机系统的计算能力提出越来越高的需求。EMAN是对冷冻电子显微镜拍摄出来的图像进行单颗粒三维重构的流行软件包。由于GPU显示出优于CPU的性能水平,在GPU机群系统上对EMAN进行加速将有可能取得较好的效果。首先,有必要调查EMAN在计算方面的特点,包括计算密度,访存情况和并行潜力。结果显示其在GPU体系结构上有获得较高性能的潜力。然后,根据此特点完成一款并行EMAN软件,使用CUDA和MPI对EMAN实施不同粒度的并行处理。在一个拥有16颗Intel6核处理器的机群系统上,用8颗FermiGPU加速后的EMAN(CUDA-EMAN)相比于仅使用CPU的并行EMAN程序获得了2.5倍的加速比,也就是1颗FermiGPU相比一个Intel6核处理器可对EMAN加速5倍。Recently, high-resolution 3D reconstruction from cryo-electron microscopy （cryo-EM） images requires more and more computational capability on current computer systems. EMAN is a popular toolkit used for single-particle 3D recon-struction of biological molecules. As GPU shows performance advantage over CPU, accelerating EMAN on a GPU cluster system is a good choice. Firstly, it is necessary to investigate EMAN’ s computing features including computational inten-sity, memory and parallelism. The result shows the potential to achieve high performance on GPU architecture. Secondly, the parallelized EMAN software that exploits multi-grain parallelism using CUDA and MPI, is presented. Comparing to CPU-only parallelized EMAN on a cluster based on sixteen Intel six-core processors, the accelerated EMAN （CUDA-EMAN） on an eight Fermi GPUs cluster accelerates the program by about 2.5 times. That means the speedup we ob-tained on one Fermi GPU is about 5x versus one Intel six-core processor.

关 键 词：EMAN GPU CUDA 并行 三维重构 

分 类 号：TP319[自动化与计算机技术—计算机软件与理论]