一种并行计算通信优化策略

A strategy of parallel computerization for optimizing communication

为了解决在ABEEMσπ模型电荷分布计算中,基于Cholesky分解并行算法在通信过程中不断增长引起的通信开销和同步开销问题,采用linux并行计算集群系统层通信优化方法,设计并实现了系统层零拷贝通讯优化方法,以及可卸载的LKM模块机制缩短开发和测试时间,减少数据拷贝冗余的基于地址映射的零拷贝技术,解决了通讯延迟等问题。测试结果表明:基于零拷贝的通讯优化方法保证了数据信息的及时性、高效性与准确性,大大提高了Cholesky分解并行算法的通信性能。

To solve the problem of the communication overhead and synchronous overhead increase in electron charge distribution calculation of ABEEM model,this study applies Linux parallel computing system communication to develop a system layer zero copy communication optimization method based on Cholesky decomposition parallel algorithm in communication process.In addition,a LKM module,which can be removed,is developed to shorten development time and test time,and reduce data copies based on the address mapping zero copy technology.It has solved the problem of communication delay.The study results show that the communication optimization method developed based on zero copy technology can achieve a timely,efficiently and accurately data communication and greatly improve the performance of Cholesy decomposing parallel algorithm.