摘要：Recently, several wireless/optical datacenter architectures are designed to overcome the drawbacks of wired datacenter topologies, such as expensive high-end switches, high cabling complexity, congestion caused by a few hot nodes. Compared with wired switches, current commodity wireless switches usually suffer lower throughput as well as higher packet loss ratio and latency. However, todays datacenter applications, such as web search, require quite low latency to improve user experience. In this paper, a mechanism, named pECN, is proposed to achieve low latency for delay-sensitive flows. pECN makes use of Explicit Congestion Notification (ECN) to maintain a small queue length and a priority-based scheduling mechanism to provide delay differentiated services. The priority-based scheduling mechanism works at the packet level. When transmitting, the packet with the highest priority will be sent. Since the queue length is controlled within a small value by the ECN mechanism, the scheduling mechanism does not incur large overhead. Also, since it works at the packet level, it can support unlimited number of services. Besides, pECN does not need to maintain any flow state at switches. We implemented the proposed mechanism and evaluated the performance on the ns-2 platform. The simulation results show that pECN works better compared with TCP, DCTCP and D2TCP in terms of the flow completion time and the number of flows that can be supported without compromising deadlines.
关键词：datacenter network; latency; ECN; priority based scheduling
