Exploring the Network-Sensitive Scheduling on Distributed Shared Memory
Xing Wei, Huiqi Hu, Aoying Zhou · 2021
Distributed shared memory abstraction can organize a cluster of machine nodes to empower the application execution with the big memory spaces and abundant parallelism. However, when deploying an application on the distributed shared memory, these loosely-coupled machine nodes may also make the execution of its tasks suffer from the frequent network traffic based on the existing scheduling strategies. In this paper, we conduct a detailed case study to reveal that the network-intensive task and poor data locality contribute to the major part of network traffic, which could impose the network burden on the execution. To address this issue, we propose the network-sensitive scheduling that fully takes the network traffic into the consideration of scheduling on DSM. The experimental results reveal that this network-sensitive scheduling can reduce the execution time by up to 32%, and decrease the network traffic by 18% under the real-world cases.