ExpressPass++: Credit-Effecient Congestion Control for Data Centers
Zejia Zhou, Dezun Dong, Shan Guo Huang, Zihao Wei · 2019
For data center networks (DCNs), the network fabric evolves a lot to serve global users, but network congestion inevitability occurs when traffic bursts as well. To efficient manage congestion in current high-speed DCNs, researchers devote much attention to the proactive congestion control which has the merits of fast convergence, low latency, and low buffer occupancy. ExpressPass, as a state-of-the-art proactive scheme, leverages credit reservation mechanism to achieve fast transmission, where each credit schedules one maximum transmission unit (MTU) data packet. However, for workloads with a high volume of very short flows that shorter than an MTU, the credit efficiency is degraded and the link bandwidth is wasted. This paper proposes ExpressPass++ to improve the credit scheduling mechanism of ExpressPass. ExpressPass++ develops a token-scheduled protocol, each credit corresponds to an MTU size tokens which can schedule an MTU amount of data. Since a credit may schedule multiple small packets, data packets will carry reduplicative credit sequence number that used for feedback control and the credit rate will be inaccurately regulated. To address this problem, ExpressPass++ improves the sequence-based feedback control of ExpressPass. Compared to ExpressPass, the results show that ExpressPass++ greatly improves in terms of average FCT and network utilization, and ExpressPass++ maintains the ExpressPass's advantages of the bounded queue and fast convergence.