计算机科学与探索
計算機科學與探索
계산궤과학여탐색
JOURNAL OF FRONTIERS OF COMPUTER SCIENCE & TECHNOLOGY
2015年
9期
1093-1099
,共7页
方宝辉%徐金秀%魏敏%周明忠
方寶輝%徐金秀%魏敏%週明忠
방보휘%서금수%위민%주명충
气候模式T106%MIC架构%混合异构编程%Offload模式
氣候模式T106%MIC架構%混閤異構編程%Offload模式
기후모식T106%MIC가구%혼합이구편정%Offload모식
BCC_AGCM_T106%MIC architecture%hybrid programming%Offload pattern
气象数值模式是天气预报和气候预测的基本工具和方法,随着技术的发展,模式分辨率有了大幅的提高,分辨率的提升使得计算量呈指数倍的增大,然而气候气象预报的时效性对并行程序的设计与计算平台的性能都提出了更高的要求。以气候模式T106为研究案例,以Intel? Xeon? PhiTM为实验平台,探索混合异构编程与优化的可行性,实现了CPU端MPI(message passing interface)+MIC(many integrated core)端OpenMP的混合异构编程,充分继承原始代码的MPI级并行,节约了开发成本。以两个CPU进程和一块MIC卡为例来测试性能数据,结果显示随着MIC卡上的线程数增多,气候模式T106核心段在MIC上加速明显,但相对于未使用MIC的纯MPI程序加速效果并不明显,这主要是由于T106核心段计算量不足而MIC卡与主机端数据交换较多造成的。
氣象數值模式是天氣預報和氣候預測的基本工具和方法,隨著技術的髮展,模式分辨率有瞭大幅的提高,分辨率的提升使得計算量呈指數倍的增大,然而氣候氣象預報的時效性對併行程序的設計與計算平檯的性能都提齣瞭更高的要求。以氣候模式T106為研究案例,以Intel? Xeon? PhiTM為實驗平檯,探索混閤異構編程與優化的可行性,實現瞭CPU耑MPI(message passing interface)+MIC(many integrated core)耑OpenMP的混閤異構編程,充分繼承原始代碼的MPI級併行,節約瞭開髮成本。以兩箇CPU進程和一塊MIC卡為例來測試性能數據,結果顯示隨著MIC卡上的線程數增多,氣候模式T106覈心段在MIC上加速明顯,但相對于未使用MIC的純MPI程序加速效果併不明顯,這主要是由于T106覈心段計算量不足而MIC卡與主機耑數據交換較多造成的。
기상수치모식시천기예보화기후예측적기본공구화방법,수착기술적발전,모식분변솔유료대폭적제고,분변솔적제승사득계산량정지수배적증대,연이기후기상예보적시효성대병행정서적설계여계산평태적성능도제출료경고적요구。이기후모식T106위연구안례,이Intel? Xeon? PhiTM위실험평태,탐색혼합이구편정여우화적가행성,실현료CPU단MPI(message passing interface)+MIC(many integrated core)단OpenMP적혼합이구편정,충분계승원시대마적MPI급병행,절약료개발성본。이량개CPU진정화일괴MIC잡위례래측시성능수거,결과현시수착MIC잡상적선정수증다,기후모식T106핵심단재MIC상가속명현,단상대우미사용MIC적순MPI정서가속효과병불명현,저주요시유우T106핵심단계산량불족이MIC잡여주궤단수거교환교다조성적。
Numerical weather model is a basic method and tool of weather forecasting and climate prediction. With the development of technology, the model resolution has improved greatly which brings out the exponentially-increasing computation cost. The time effectiveness of weather forecasting and climate prediction puts forward more advanced requests for the design of parallel program and the performance of computing platform. This paper takes BCC_AGCM_T106 for example to explore the feasibility of hybrid programming and optimization on the Intel? Xeon? PhiTM, implements the hybrid programming of MPI (message passing interface) parallel computation on CPU and OpenMP parallel computation on MIC (many integrated core). The experimental results show that the method can inherit the most of the original MPI parallel computing codes and reduce the developing costs significantly. This paper gets the test data using two CPU processes and an MIC card as an example. The results show that the acceleration performance of climate model T106 core section in MIC accelerates obviously with the number of threads on the MIC card increasing. But compared to the pure MPI program without using the MIC card, acceleration effect is not obvious, which is mainly due to that the calculation amount of T106 core section is insufficient and the data exchange between the MIC card and the host side is too much.