A Buffer Management Algorithm Based on Dynamic Marking Threshold to Restrain MicroBurst in Data Center Network

• Main Authors: Yan Yu, Xianliang Jiang, Guang Jin, Zihang Gao, Penghui Li
• Format: Article
• Language: English
• Published: MDPI AG 2021-09-01
• Series: Information
• Subjects: data center network; flow scheduling; active queue management; Incast; explicit congestion notification
• Online Access: https://www.mdpi.com/2078-2489/12/9/369

The data center has become the infrastructure of most Internet services, and its network carries different types of business flow, such as query, data backup, control information, etc. At the same time, the throughput-sensitive large flows occupy a lot of bandwidth, resulting in the small flow’s longer completion time, finally affecting the performance of the applications. Recent proposals consider only dynamically adjusting the ECN threshold or reversing the ECN packet priority. This paper combines these two improvements and presents the HDCQ method for coordinating data center queuing, separating large and small flows, and scheduling in order to ensure flow completion time. It uses the ECN mechanism to design load-adaptive marking threshold update algorithms for small flows to prevent micro-bursts from occurring. At the same time, packets marked with ECN or ACK are raised in priority, prompting these packets to be fed back to the sender as soon as possible, effectively reducing the TCP control loop delay. Extensive experimental analysis on the network simulator (NS-2) shows that the HDCQ algorithm has better performance in the face of micro-burst traffic, reducing the average flow completion time by up to 24% compared with the PIAS.