摘要
在通信过程中,不同计算节点会运行同一个程序,每个节点拥有用来区分不同功能的唯一编号Rank,在此过程中,部分节点会频繁的通信,进而产生不同的跳数。为了加快数据交换速度,使得全局的计算代价最小化,根据现有节点间时延测试数据记录、不同Rank之间的通信频次等数据,使用最小二乘法拟合出时延和跳数的关系,并利用整数规划以及Dijkstra算法将Rank分配到各节点,得到合理的任务映射策略,经过验证比当前方法有所优化。
In the communication process,different compute nodes will run the same program,and each node has a unique number Rank used to distinguish different functions.In this process,some nodes will communicate frequently,and thus produce different hops.In order to speed up data exchange and minimize the global computational costs,we use least squares method to fit the relationship between delay and hop count based on the existing data such as inter-node latency test data records and the frequency of communication between different Ranks,and use integer programming and Dijkstra algorithm to assign the Ranks to each node to obtain a reasonable task mapping strategy.It is verified to be optimized over current methods.
作者
王楚越
文万志
Wang Chuyue;Wen Wanzhi(School of Information Science and Technology,Nantong University,Nantong,Jiangsu 226000,China)
出处
《计算机时代》
2022年第12期13-16,共4页
Computer Era
基金
南通市基础科学研究项目(JC2021125、JCZ21087)。