基于线性阵列处理器的GRAPES核心代码优化  

GRAPES Kernel Acceleration on Linear Function Array Based Processor

作　　者：王为[1] 张悠慧[1] 姚骏 李艳华[1] 郑纬民[1] 

机构地区：[1]清华大学计算机科学与技术系,北京100084 [2]奈良先端科技大学院大学信息科学学院

出　　处：《计算机学报》2013年第10期2053-2061,共9页Chinese Journal of Computers

基　　金：日本东京大学VLSI设计与教育中心(VDEC);Synopsys and Cadence Corporation;JST ALCA;KAKENHI(Nos.24240005;24650020;and 2370060);日本JST A-STEP与日本奈良科技大学院大学全球创新创新计划基金(AS232Z02313A);国家"八六三"高技术研究发展计划项目基金(2013AA01A215)资助~~

摘　　要：我国气象局研究开发的数值天气预报系统GRAPES作为典型的高性能计算应用,在人民日常生活中有着极其重要的作用,如何提高GRAPES系统性能并控制其功耗以支持因增加系统分辨率而急剧增加的运算量是一个重大课题.该文使用软硬结合的多种方法对GRAPES系统的核心代码进行优化.采用线性阵列流水处理器LAPP实现循环级并行;采用循环切分、数据预取、缓存分区、多路预取等方法来进行加速;采用电源门控等低功耗技术来降低功耗.实验结果表明,优化后的GRAPES核心模块运行时平均IPC可以达到11.3,是面积相同的通用多核处理器的2.3倍;低功耗技术使其功耗仅为通用多核处理器的12%;同时优化后的LAPP集群性能功耗比可以达到相同计算能力Intel Xeon集群的11.7倍.With the continuous scaling of the semiconductor process technology, performance improvement simply by increase the frequency is not feasible any more. Parallelism has become the key technology to obtain energy-efficiency in large scale applications. How to keep improving performance with the energy consuming under control has become a key problem. In this paper, a high performance computing weather forecasting application was tuned using a functional unit （FU） array based architecture. Software approaches like Array reforming, Loop rescheduling, local buffer pre-fetching, local buffer partition, and hardware approaches like pre-fetching paralleliza- tion are specifically employed to increase the data locality and data reuse to accelerate the stencil computation in GRAPES. The simulation results indicate that we can have an average IPC of 11.3, which is 2.3x of the multi-core processors, while only consuming 12% of its energy. This can accordingly reduce communication in the cluster system, resulting in a 11.7x power-effcieney boost.

关 键 词：计算机系统结构 数值天气预报 赫姆霍兹方程 Stencil计算 

分 类 号：TP303[自动化与计算机技术—计算机系统结构]