GRAPES有限区域伴随模式内存优化  

Memory Optimization of the Adjoint Model of the GRAPES Regional Model

作　　者：李金才[1] 赵文涛[1] 赵军[1] 

机构地区：[1]国防科学技术大学计算机学院,湖南长沙410073

出　　处：《计算机工程与科学》2012年第6期65-69,共5页Computer Engineering & Science

基　　金：国家自然科学基金资助项目(40775064);中国气象局武汉暴雨研究所2008年度暴雨研究开放基金资助项目(IHR2008K04)

摘　　要：GRAPES有限区域伴随模式是基于自动微分工具TAPENADE转换与手工代码编写相结合的方式开发而成,主要由模式程序和内存支撑函数库(PLP库)构成。由于TAPENADE采用全存储策略来保存中间变量,造成了伴随模式运行过程中内存开销过大,并且出现随着时间步的增加内存不断增长的情况。对伴随模式内存支撑函数库中POP类函数算法进行修改,解决了内存增长的问题;从整型变量、实型变量和正模式子程序调用三个方面对模式程序进行优化,使得伴随模式运行时的内存开销显著减少。The adjoint model of the GRAPES regional model has been developed with a combination of automatic differentiation tool TAPENADE and hand-coding. The model is made up of mode programs and a memory supporting library （PLP library）. The TAPENADE employs a complete storage strategy to save the intermediate variables. This causes too much memory overhead during the operation of the adjoint mode and increases memory consumption with the time step going on. The growing of memory is addressed by modifying the POP functions in the PLP. The mode programs are optimized for integer variables, real variables and calling subroutines of the forward model so that the memory overhead is significantly decreased.

关 键 词：数值天气预报 GRAPES 伴随模式 内存优化 TAPENADE 

分 类 号：TP301.1[自动化与计算机技术—计算机系统结构]