1. Scalable Congestion Control Protocol based on SDN in Data Center Networks Speaker : Bo-Han Hua Professor : Dr. Kai-Wei Ke Date : 2016/04/08 1

2. Outline Introduction Software Defined Network (SDN) SCCP Simulation Conclusion Reference 2

3. Introduction Data centers have emerged as a key resource in the past decade for providing a plethora of online services such as web search and cloud computing. 3

4. Introduction(Cont.) It proposes a new scalable congestion control protocol, called SCCP, that is implemented to the Software Defined Networking (SDN) data center switches. Its main objective is to avoid the switch buffer overflow and reduce the queuing delay even under the circumstances where a massive number of bursty flows. 4

5. SDN 5

6. SDN (Cont.) 6

7. 7

8. SCCP Scalable Congestion Control Protocol Extend the OpenFlow specifications at the SDN switch. Accurately assess the number of TCP flows that traverse each switch port and be used to compute the fair-share of each flow. Let the total link utilization does not exceed the bandwidth delay product (BDP). 8

9. SCCP (Cont.) Each port is shared by multiple flows. The output-port is where the network congestion in data centers mainly occurs. So its congestion control scheme is conducted at each switch output-port. 9

10. SCCP (Cont.) 1. For an outgoing packet : Keep track of the number of TCP flows (N i ) 2. For an incoming packet : Compute the fair-share(fair_share i ) Change the value of the advertisement window field (awnd) in TCP header, to the value of fair_share i, if awnd > fair_share i 10

11. SCCP (Cont.) 11 Compute the fair-share by simply dividing the BDP of the ith port by N i. In the BDP, the bandwidth is the link capacity of the port, and the delay is the common RTT of all flows. So the fair-share is computed as follows:

12. SCCP (Cont.) To support SCCP operations in the SDN framework, we need to extend the flow match fields defined in the OpenFlow specification. (version 1.5.0) 12

13. SCCP (Cont.) OFPXMT_OFB_TCP_WINDOW The fair-share information is set via the set-window action. OFPAT_INCREMENT_NI Increase N i by one when the port number i is given. OFPAT_DECREMENT_NI Decrease N i by one when the port number i is given. OFPAT_SET_ WINDOW Set a header field if the new value is smaller than the existing value. 13

14. SCCP (Cont.) 14 SCCP fair_share example

15. Simulation Implement the SCCP algorithm in the ns-3 simulator. It consists of 1 core, 2 aggregation, 10 Top- of-Rack (ToR) switches, and 48 servers per rack. Link Rate : 1Gbps for ToR switches 10Gbps for core/aggregation switches 15

16. Simulation(Cont.) Packet buffer size per port : 128 Kbytes for ToR 300 Kbytes for aggregation 400 Kbytes for core switches Make comparisons with 1. NewReno + Droptail 2. DCTCP + ECN 3. NewReno + SCCP. 16

17. Simulation(Cont.) 17 Simulation topology

18. Simulation(Cont.) Single root scenario 1 Root & n workers (n:10~400) Response data size of each worker is 1MB/n bytes. One background flow that transmits 10MB of data to the root as the median number of concurrent large flows is 1 in data center networks. The background flow fully utilizes the bottleneck link before the Partition/Aggregate application begins. 18

19. Simulation(Cont.) 19 Single root scenario (Cont.)

20. Simulation(Cont.) 20 Single root scenario (Cont.)

21. Simulation(Cont.) Multiple root scenario Deploy multiple Partition/Aggregate applications to see how the protocols work when the network congestion occurs at the aggregation/core switches as well. Multiple roots are evenly distributed over the first 5 racks, and each root has 90 workers. The transmitting data size of each worker is 20KB. 21

22. Simulation(Cont.) Multiple root scenario 22

23. Simulation(Cont.) Multiple root scenario (Cont.) 23

24. Simulation(Cont.) Multiple root scenario (Cont.) 24

25. Conclusion To minimize service latencies,and provide better quality of service in the data centers, it is essential to fundamentally address the network congestion, caused by the momentary bursts traffic. To deal with a large number of bursty flows, it designs SCCP to provide more exact information, the fair-share of each flow, to the senders quickly via the advertisement window field in the TCP header. 25

26. Reference Jaehyun Hwang, Joon Yoo, & Sang-Hun Lee (2015). Scalable Congestion Control Protocol Based on SDN in Data Center Networks. 2015 IEEE Global Communications Conference (GLOBECOM), 1-6. Kreutz, D., Ramos, F. M., Esteves Verissimo, P., Esteve Rothenberg, C., Azodolmolky, S., & Uhlig, S. (2015). Software-defined networking: A comprehensive survey. Proceedings of the IEEE, 103(1), 14-76. 26

27. Thanks for listening. 27