Towards Lightweight Traffic Forecasting in RDMA Networks: Design and Application

Communication becomes the bottleneck of data-parallel computing systems. Although Remote Direct Memory Access (RDMA) was proposed to solve the communication bottleneck at end hosts, network congestion can still prolong data transmission time, thereby degrading application performance. To avoid network congestion, we need to deploy bandwidth provisioning or traffic engineering schemes, both of which require ahead-of-time traffic information as input. In this work, we design ApOLLO to accurately forecast the traffic in RDMA networks. The key idea of ApOLLO is to estimate the traffic amount that will be injected into RDMA networks based on the information recorded in memory or RDMA NICs. To reduce the system overhead and achieve timely forecasting, ApOLLO uses shared memory to transfer forecasting results among its processes. By implementing ApOLLO via modifying RDMA Verbs APIs, ApOLLO is ready-to-deploy and transparent to users. Through extensive experiments on a real testbed, we demonstrate that ApOLLO can achieve accurate traffic forecasting with less than 10% CPU usage and can help reduce the average flow completion time by up to 28.9% when combined with a naive load balancer.