结合通信重排和消息合并的通信调度方法研究

网络通信对于高性能计算机应用至关重要。当前,随着数值模拟应用的复杂化和并行规模的不断提升,应用软件对于缓解拥塞和减少通信协议开销的需求愈发迫切。传统的消息合并方法只以减少通信协议开销和延迟为目标,所以针对小消息进行合并。与之不同的是,从调度算法的角度提出了一种通过消息重排以减缓大消息网络拥塞,并基于优先级合并消息来提高网络有效利用率的算法。实验表明,该算法针对真实应用的通信性能最大可以提升41%,平均对每个应用提升了10%。

A communication scheduling method combining communication rearrangement and message merging

Abstract: Network communication is critical for high-performance computer applications. At present, with the complication of numerical simulation applications and the increasing scale of parallelism, the need for application software to alleviate congestion and reduce communication protocol overhead is becoming more and more urgent. The traditional message merging method only merges small messages with the goal of reducing the communication protocol overhead and latency. In contrast, from the perspective of scheduling algorithms, this paper proposes an algorithm for reducing the network congestion of large messages through message rearrangement and improving the effective utilization of the network by merging messages based on priority. Experiments show that our algorithm can increase the communication performance of real applications by up to 41%, and on average by 10% for each application.