Flow Consolidation for Congestion Control in Data Centers

Congestion control has been a fundamental component with evolving challenges in computer networks during the past few decades. In this thesis, we focus on the following key challenges of congestion control: lack of information sharing among flows, limited congestion visibility at the end-hosts, stringent requirements for convergence time, and dynamic and volatile nature of networks. We design a hierarchical congestion control system (HCC) that addresses the above challenges in data center networks. By grouping flows in a hierarchical manner, HCC enables flow cooperation and information sharing, leading to expanded visibility for flows, improved fairness, and faster convergence. Flows share congestion information within groups and propagate the compressed signals through the hierarchy, enabling aggregate control of multiple flows. HCC implements a distributed system for realizing max-min fairness, which is the theoretical objective of many congestion control protocols. To limit the state space, communication overhead, and processing time, HCC leverages flow aggregation and rate quantization techniques. Rate quantization is an effective way to reduce the run-time of max-min fairness. Flow aggregation helps flows to react collaboratively to changes at the edge of the network and converge to a local optimal state promptly. Also, the aggregated updates are propagated through the hierarchy to reach a globalmax-min fair state. This fast reaction at the edge is beneficial for short-lived flows and on-off traffic patterns. Moreover, to eliminate the volatile nature of flows, we define the correlation-aware flow consolidation problem. In correlation-aware flow consolidation, the goal is to reduce the fluctuations in individual flows by aggregating inversely correlated (or uncorrelated) flows. In our experiments, we show that we can reduce the average standard deviation of group demands by 33% which results in estimating future group demands with higher confidence. We evaluate HCC on a real workload and show that HCC converges to full utilization much faster (up to 4x), with a near zero bottleneck queue size, lower flow completion times (42 − 62%), and a significantly higher fairness index compared to well-known congestion control protocols. Also, by comparing HCC with a centralized solution, we show that HCC significantly reduces the overheads (by 73% and 99.5%).

Department dc:contributor.department

Computer Science

Year dc:date.issued

2023

Author dc:creator

• Ketabi, Shiva

Advisor dc:contributor.advisor

• Ganjali, Yashar

Handle dc:identifier.uri

http://hdl.handle.net/1807/128252

OAI identifier oai:identifier

oai:utoronto.scholaris.ca:1807/128252