CWND: A Coarse, But Simple, Efficient Metric to Reduce Short Flow Completion Time in Data Centers

Reducing the delay-sensitive Short Flows Completion Time (SFCT) is a hot topic in data center networks, and prioritizing the short flows is an effective way to do that. However, precisely identifying a short flow is still a challenge, which makes many proposed schemes unfeasible in commodity switch. The short flows often have a small Congestion Window (CWND) due to their small flow sizes, and thus CWND can be used as a coarse metric to identify short flows. Based on this idea, Low Delay Data Center Transport Control Protocol (LD-DCTCP) is proposed. When transferring a data packet whose CWND is less than the CWND threshold, LD-DCTCP sender marks the high priority field in the IP header of the packet. Then, switches can prioritize the high priority packets by a priority-based queue schedule. By simply prioritizing the packets which have small CWND, short flows, in a high probability, can be prioritized, and as well, SFCT can be effectively reduced. Since CWND can be easily obtained from TCP/IP header, LD-DCTCP is implementable in data center networks. The NS2 simulation results show that the average flow completion time of short flows in DCTCP can be reduced by 26%-40% in normal cases. Furthermore, due to the coarseness of CWND metric, some cases in which CWND-based judgment could be wrong are also discussed in the paper, and the simulation results show that no severe problem occurs after using LD-DCTCP.

[1]  Alex X. Liu,et al.  Friends, not foes: synthesizing existing transport strategies for data center networks , 2015, SIGCOMM 2015.

[2]  John K. Ousterhout,et al.  Homa: a receiver-driven low-latency transport protocol using network priorities , 2018, SIGCOMM.

[3]  Nick McKeown,et al.  pFabric: minimal near-optimal datacenter transport , 2013, SIGCOMM.

[4]  Brighten Godfrey,et al.  Finishing flows quickly with preemptive scheduling , 2012, CCRV.

[5]  David A. Maltz,et al.  DCTCP: Efficient Packet Transport for the Commoditized Data Center , 2010 .

[6]  Christo Wilson,et al.  Better never than late , 2011, SIGCOMM 2011.

[7]  Ali Munir,et al.  Minimizing flow completion times in data centers , 2013, 2013 Proceedings IEEE INFOCOM.

[8]  Wei Bai,et al.  Information-Agnostic Flow Scheduling for Commodity Data Centers , 2015, NSDI.

[9]  Gautam Kumar,et al.  pHost: distributed near-optimal datacenter transport over commodity network fabric , 2015, CoNEXT.

[10]  T. N. Vijaykumar,et al.  Deadline-aware datacenter tcp (D2TCP) , 2012, CCRV.

[11]  Ying Zhang,et al.  Providing bandwidth guarantees, work conservation and low latency simultaneously in the cloud , 2016, IEEE INFOCOM 2016 - The 35th Annual IEEE International Conference on Computer Communications.

[12]  Mark Handley,et al.  Re-architecting datacenter networks and stacks for low latency and high performance , 2017, SIGCOMM.

[13]  Amiya Nayak,et al.  Improving flow completion time for short flows in datacenter networks , 2015, 2015 IFIP/IEEE International Symposium on Integrated Network Management (IM).

[14]  Devavrat Shah,et al.  Fastpass: a centralized "zero-queue" datacenter network , 2015, SIGCOMM 2015.

[15]  Amin Vahdat,et al.  Less Is More: Trading a Little Bandwidth for Ultra-Low Latency in the Data Center , 2012, NSDI.