Intelligent queue management of open vSwitch in multi-tenant data center

Multi-tenant data centers (MTDCs) host numerous applications with dominant transport layer protocol TCP, hence the performance of TCP matters a lot. However, it is difficult for the tenants to configure their TCP/IP stacks appropriately due to the performance optimization for TCP is complicated on account of numerous parameters to be tuned manually according to different scenarios. Furthermore, the multiple tenants may interfere with each other, which leads to unfairness and performance degradation. For the cloud service provider, who cannot touch the tenants' stacks, it is a formidable challenge to achieve consistent and advanced network performance in MTDC. This paper proposes a scheme to improve network performance by deploying a DRL-based queue management algorithm (DRLQM) in the virtual switch (vSwitch), which is transparent, pluggable, and zero-configuration. We construct a training simulator, which supports different objectives (low delay and high throughput) balanced and converged, and then implement a prototype in the Linux kernel with the fine-trained queue management model. Various experiments show that DRLQM achieves remarkable performance improvements, in terms of throughput, Round-Trip Time (RTT), loss ratio, and fairness, both in simulation and in the real network. Generally, DRLQM can reduce the retransmission by 86%, and 92%, and the RTT by 62%, and 87% compared to First-in-First-out queue (FIFO) and multi-queue (MQ), respectively.(c) 2023 Elsevier B.V. All rights reserved.

Automated summary

This paper proposes a scheme to improve network performance in multi-tenant data centers (MTDCs) by deploying a DRL-based queue management algorithm (DRLQM) in the virtual switch (vSwitch). Various experiments show that DRLQM achieves remarkable performance improvements in terms of throughput, RTT, loss ratio, and fairness.