UnisonFlow: A Software-Defined Coordination Mechanism for Message-Passing Communication and Computation
UnisonFlow: A Software-Defined Coordination Mechanism for Message-Passing Communication and Computation
复制标题
UnisonFlow：用于消息传递通信和计算的软件定义的协调机制
DOI：
10.1109/access.2018.2829532
复制
发表时间：
2018
期刊：
影响因子：
3.9
通讯作者：
Shinji Shimojo
中科院分区：
文献类型：
--
作者：
Keichi Takahashi;Susumu Date;Dashdavaa Khureltulga;Yoshiyuki Kido;Hiroaki Yamanaka;Eiji Kawai;Shinji Shimojo
消息传递接口（MPI）通信性能正成为影响集群数据密集型应用程序整体性能的关键因素之一。我们的软件定义网络（SDN）增强MPI通过将SDN的灵活和动态网络可控性集成到MPI中来提高互连通信的性能。我们已经证明，个人MPI通信原语的加速是可行的，通过我们过去的工作SDN增强MPI。然而，现实世界的MPI应用程序并没有受益于这样的加速通信原语通过我们的研究成果，到目前为止，因为每个不同的网络控制算法设计的各种MPI通信原语不能被激活和协调的MPI应用程序的执行。因此，本文提出了UnisonFlow，一个软件定义的协调机制，SDN增强的MPI，执行网络控制与应用程序的执行同步。在实际计算机集群上进行的实验验证了互连控制可以成功地与应用程序的执行同步执行。此外，同步以低开销执行，并且其性能损失实际上可以忽略不计。
Message passing interface (MPI) communication performance is becoming one of the key factors heavily affecting the total performance of data-intensive applications running on computer clusters. Our software-defined networking (SDN)-enhanced MPI improves the performance of communication over interconnects by integrating flexible and dynamic network controllability of SDN into MPI. We have demonstrated that the acceleration of individual MPI communication primitives is feasible through our past work on the SDN-enhanced MPI. However, real-world MPI applications have not benefited from such accelerated communication primitives through our research achievements to date, because each of the distinct network control algorithms designed for various MPI communication primitives cannot be activated and coordinated with the execution of the MPI application. Therefore, this paper proposes UnisonFlow, a software-defined coordination mechanism for the SDN-enhanced MPI that performs network control in synchronization with the execution of applications. An experiment conducted on a real-computer cluster verifies that the interconnect control can be successfully performed in synchronization with the execution of the application. Furthermore, the synchronization is performed with a low overhead and its performance penalty is practically negligible.