摘要
对基因表达谱分块,使之符合GPU并行计算的线程结构特性,根据GPU线程结构特性设计双层并行模式,并利用纹理缓存实现访存高效;依据CPU二级缓存容量对基本块进一步细分成子块以提高缓存命中率,利用数据预取技术减少访存次数,利用线程绑定技术减少线程在核心之间的迁移;依据多核CPU和GPU的计算能力分配CPU和GPU的基因互信息计算任务以平衡CPU与GPU的计算负载;在设计新的阈值计算算法基础上,设计实现了访存高效的构建全局基因调控网络CPU/GPU并行算法.实验结果表明,与已有算法相比,本文算法加速更明显,并且能够构建更大规模的全局基因调控网络.
Gene expression profile is parted into some basic blocks to conform GPU parallel computing threads structural characteris- tics, a two-level GPU parallel processing mode is designed by using the characteristic of thread structure for GPU ,and efficient memo- ry access is achieved by using texture cache memory. Each basic block is further divided to several sub-blocks according to the capaci- ty of L2 cache for CPU in order to enhance the hit rate of access cache, the data prefetching techniques are used to reduce memory ac- cess times, the thread-bound techniques are used to reduce migration between threads in the core. The tasks to compute mutual infor- mation between genes are distributed to CPU and GPUs according to their computation ability in order to balance the computation loads for CPU and GPU. Based on designing a new threshold computation algorithm, this paper proposes the cache-efficient CPU/GPU parallel algorithms for constructing genome-wide gene regulatory networks. Experimental results show that, compared with the existing parallel algorithm, our algorithms achieve higher speedup and can construct larger scale genome-wide gene regulatory networks.
出处
《小型微型计算机系统》
CSCD
北大核心
2015年第2期234-239,共6页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(61462005)资助
广西自然科学基金项目(2014GXSFAA118396)资助
广西教育厅-广西大学博士点建设基金项目(P11900119)资助
广西研究生教育创新计划项目(YCSZ2013006)资助
关键词
全局基因调控网络
CPU与GPU协同计算
访存高效
并行算法
genome-wide gene regulatory network
CPU/GPU cooperative computing
efficient access cache
parallel algorithm
