An urgency and congestion control scheme for larger-scale TCP incast problem in data center

Cloud computing has been widely used to provide services such as search, social networks, and distributed storage systems. Data centers are built to support the requests of cloud applications which generally adopt multi-layer partition/aggregation workflow pattern. In this multi-hop environment with dual-homed topology, the aggregator needs to wait for all responses from the works to generate final results. We observe that the incast congestion is inevitable as the number of workers or aggregators increases. In addition, other challenge is to meet soft real-time constraints under the SLA at each aggregator. Prior works proposed either deadline-agnostic protocol or deadline-aware protocol without considering other congested flows. In this paper, we propose an urgency and congestion control scheme (UCCS) that detects the network congestion situation by considering urgency and congestion degree of the transmitted flows on each path and then adjusts the transmission rates. The simulation results show that UCCS can effectively improve the TCP incast problem, network throughput, flow completion time, and deadline guarantee.