【摘要】：神威太湖之光是最新一期Top500榜單上排名第一的超級計算機,峰值性能為125.4PFlops,其計算能力主要歸功于國產SW26010眾核處理器。OpenFOAM(Open Source Field Operation and Manipulation)是計算流體力學領域使用最廣泛的開源軟件包,但是由于其基于C++實現(xiàn),與神威太湖之光上的異構眾核處理器SW26010的編譯器不兼容,因此無法直接在該架構上有效運行。基于SW26010的主核/從核的體系架構移植了OpenFOAM的核心計算代碼,并采用混合語言編程實現(xiàn)的方式來解決編譯不兼容的問題。此外,通過寄存器通信、向量化和雙緩沖等優(yōu)化手段,單核組的性能較優(yōu)化后的主核代碼提高了8.03倍,較Intel(R)Xeon(R)CPU E5-2695v3的串行執(zhí)行性能提高了1.18倍。同時,將單核組的實現(xiàn)擴展到了神威太湖之光的大規(guī)模集群上,并進行了強可擴展性測試,256個核組上實現(xiàn)了184.9倍的加速。采用的移植方式和優(yōu)化手段也可以為其他復雜C++程序在神威太湖之光上的應用提供借鑒。
[Abstract]:Shenwei Taihu Light is the No. 1 supercomputer on the latest Top500 list, with peak performance of 125.4 PFlops.The computational power is mainly attributed to the domestic SW26010 multicore processor. OpenFOAM (Open Source Field Operation and Manipulation) is the most widely used open source software package in computational fluid dynamics. However, because it is based on C, and is incompatible with the compiler of SW26010, a heterogeneous multi-core processor on the light of Taihu Lake, it can not run effectively on this architecture directly. The core computing code of OpenFOAM is transplanted based on the architecture of master / slave core of SW26010, and the problem of compiling incompatibility is solved by mixed language programming. In addition, by means of register communication, vectorization and double buffering, the performance of the single core group is 8.03 times higher than that of the optimized main core code, and 1.18 times higher than that of the serial execution performance of Intel (R) Xeon (R) CPU E5-2695v3. At the same time, the implementation of mononuclear group is extended to the large-scale cluster of light in Shenwei Lake Taihu, and a strong extensibility test is carried out, and the acceleration of 184.9 times is achieved on 256 core groups. The transplant method and optimization method can also be used for reference in the application of other complex C programs in the light of Shenwei Lake Taihu.
【作者單位】： 上海交通大學高性能計算中心;東京工業(yè)大學;
【基金】：國家重點研發(fā)計劃(2016YFB0201400,2016YFB0201800) 日本學術振興會JSPS的RONPAKU項目 并行計算機工程技術研究中心資助
【分類號】：TP338
，

本文編號：2249387