Improving Flow Completion Times in Data Center Networks

Abstract

The popularity of online services such as web search, advertisement, social networking, and recommendation systems has lead to a proliferation of data center networks in the last few years. These services share a few common themes. First, they are soft real time nature. Consequently, the response time of tasks determines the quality of results, which in turn impacts oper- ators revenue. Second, data centers typically employ a Partition/Aggregate work ow pattern, in which user requests are partitioned amongst layers of workers whose results are then aggregated. The returned responses from workers are often synchronized, leading to the Incast problem. In addition, data center applications generate a mix of tra c workloads. While most ows are short (<100 KB), a few long ows can saturate the path. Consequently, such ows can lead to large queueing delays and thus a ect delay-sensitive tra c. In this work, we leverage the above characteristics to design a protocol that seeks to reduce average ow completion time (AFCT) within data cen- ters. To this end, rst, we explore the e cacy of ow size awareness for reducing ow completion times in data centers. Second, we present the de- sign, analysis and evaluation of a size aware delivery control protocol (SDC) which incorporates these features to to achieve small ow completion times while maintaining high network throughput and small queues. Third, we present evaluation of a router assisted capacity sharing (RACS) framework. Detailed performance analysis shows that RACS can improve AFCT of ap- plications signi cantly, hence improve application throughput. Performance comparison with existing schemes like rate control protocol (RCP) and TCP- SACK shows that DC network performance can be signi cantly improved by leveraging ow size information at transport layer.