Traffic Offloading of Optic/Electric Hybrid Data Center Networks (DCNs)

1. Traffic Offloading of Optic/Electric Hybrid Data Center Networks (DCNs) Jianke Lia, Tong Yea, Tony T. Leeb aShanghai Jiao Tong University, Shanghai, ChinabThe Chinese University of Hong Kong, Shenzhen, Guangdong, China July 26, 2018

2. Overview • TrafficCharacteristic ofDCNs • Optical Offloading Process • Modelling of Optical Offloading Process • Buffer Threshold Selection • Conclusion

3. Data Centers Play Important Roles • World-wide information service infrastructure Google World-wide Data Center Map Amazon Web Service’s Global Infrastructure • [1] http://datacenterfrontier.com/regional-data-center-clusters-power-amazons-cloud/ • [2] Sushant Jain et. al., “B4: Experience with a Globally-deployed Software Defined Wan”, ACM SIGCOMM, Oct. 2013, pp. 3-14.

4. Current DCN • Servers are interconnected by a set of switches • 30-50 servers are attached to a top-of-rack (ToR) switch • Electronic switches (EPS) at the core layer

5. Traffic Characteristics in DCNs • Traffic between a pair of TORs • Very small about > 90% of the time: mice flow • Very huge about < 10 % of the time: elephant flow • Mice flow is time-sensitive but elephant flow is not • [1] A. Greenberg, et. al., “VL2: a scalable and flexible data center network,” in Proc. ACM SIGCOMM conference,2009, Barcelona, Spain.

6. Hybrid Optical/Electrical DCN • Features of Optical Circuit Switch (OCS) • Can provide high-capacity lightpaths • Reconfiguration time is on the order of • OCS deals with time-insensitive Elephant flows • [1] N. Farrington, et al., “Helios: a hybrid electrical/optical switch architecture for modular data centers,” in Proc. ACM SIGCOMM, Aug. 2010, vol. 40, pp. 339–350. • [2] G. Wang, et al., “c-Through: Part-time optics in data centers,” in Proc. ACM SIGCOMM, Aug. 2010, pp. 327–338.

7. Problem: How to trigger Lightpath? • Optical bandwidth is shared by all TORs • Lightpath is setup to serve a TOR pair only if there is an elephant flow • However, accurate elephant flow detection is difficult[1-3] • Detection error → large delay or waste of lightpath resource • Inaccurate detection → waste of control plane resource • [1] M. AI-Fares, et. al., “Hedera: Dynamic How scheduling for data center networks,” in Proc. USENIX NSDI, Apr. 2010. • [2] A. R. Curtis, et. al., “Mahout: Low-overhead datacenter traffic management using end-host-based elephant detection,” in Proc. IEEE INFOCOM, Apr. 2011, pp. 1629-1637. • [3] T. Mori, et. al., “Identifying elephant Hows through periodically sampled packets,” in Proc. IMC, Oct. 2004, pp 115-120.

8. Overview • TrafficCharacteristic ofDCNs • Optical Offloading Process • Modelling of Optical Offloading Process • Buffer Threshold Selection • Conclusion

9. Optical Offloading Strategy • ToRtransmits traffic via EPSs by default • Once buffer length exceeds a preset threshold , ToR setups a lightpathto offload the traffic • Once buffer becomes empty, ToR tears down lightpath

10. How to Select ? • If is small, lightpath setup will be frequent • Number of bandwidth requests is large • More route computing • More flow table set up • If is excessively large • Offloading is not timely, and thus delay is large • What is a good ? • Lightpath setup frequency is not high • Delay should be in an acceptable region • Adaptiveto traffic burstiness

11. Overview • TrafficCharacteristic ofDCNs • Optical Offloading Process • Modelling of Optical Offloading Process • Buffer Threshold Selection • Conclusion

12. Fluid-flow Model • Traffic arrival: ON-OFF Markov fluid flow • Traffic rate at on state: • Average duration of on state: • Traffic rate at off state: • Average duration of off state: • Burstiness: • Service state: • Electrical service rate: • Optical service rate:

13. State Transition • System states : • : Arriving process in off/on state • : Traffic is transmitted by EPS/OCS

14. Differential Equations • Let be buffer occupancy • Define steady-state cumulative distribution function(CDF) of buffer occupancy • According to state transition diagram, satisfies

15. Coefficients of Differential Equations

16. Mean Delay • Using differential equation, we have and its density function • Mean delay where are determined coefficients

17. Lightpath Setup Frequency • Light path setup frequency

18. Overview • TrafficCharacteristic ofDCNs • Optical Offloading Process • Modelling of Optical Offloading Process • Buffer Threshold Selection • Conclusion

19. Performance vs.Burstiness • In real DCNs, [1] • Mean delay and lightpath setup frequency are stable when ranges from 2 to 7 • ,,, • [1] T. Benson, A. Akella, and D. A. Maltz, “Network traffic characteristics of data centers in the wild,” in 10th ACM SIGCOMM Conf. on Internet Measurement, Nov. 2010, pp. 267–280.

20. What happens under large • Large large ON (OFF) duration: • OFF: buffer occupancy • ON: queue length variationtriangular wave

21. Theorem 1 • If is sufficiently large and • and when are very close to that when

22. Tradeoff between and • ,, • ,

23. Threshold Selection • Given delay requirement, we can select buffer threshold as follows: • Application • Delay requirement of Amazon DCNs is [1] • Assume that, ,, • Our result gives • [1] G. DeCandia, et. al., “Dynamo: Amazon’s highly available key-value store,” In Proc. 21st ACM Symposium on Operating Systems Principles (SOSP), Oct. 2007

24. Conclusion • Develop an approach to offload the burst traffic in hybrid optic/electric DCNs • Derive a closed-form solutionof mean delay and light path frequency • Provide a rule to select appropriate threshold

25. Thank You!