1. «Insomnia in the Access» Or How to Curb Access Network Related Energy Consumption Marco Canini EPFL Joint work with Eduard Goma, Alberto Lopez Toledo, Nikolaos Laoutaris (Telefonica Research), Dejan Kostić (EPFL), Pablo Rodriguez (Telefonica Research), Rade Stanojević (IMDEA Networks), Pablo Yagüe Valentín (Telefonica Research)

2. Greening the ICT Datacenters Access network ? Backbone PCs & peripherals 2

3. Energy consumption of telcos 2.1 TWh 3.7 TWh 4.5 TWh 9.9 TWh 3

4. Access dominates energy consumption ACCESS Backbone/ Metro/ Transport 4

5. A typical DSL access network CORE METRO Gateway 5 USER PARTISP PART ACCESS

6. A typical DSL access network CORE METRO USER PARTISP PART ACCESS DSL Access Multiplexers (DSLAMs) Cable bundle Central Office Gateway 6

7. Distribution frames of a Central Office 7

8. Can we save energy? How? 8

9. WHY DOES THE ACCESS CONSUME SO MUCH? 9

10. Individually, they do not consume a lot #1: Huge number of devices PhotoBlackburn But collectively … 2 orders of magnitude more gateways than DSLAMs 1 order of magnitude more DSLAMs than metro devices 2 orders of magnitude more DSLAMs than backbone devices 10

11. 2#: High per bit energy consumption At full load, access devices 2-3 orders of magnitude higher than metro/backbone ACCESS Backbone/ Metro 11

12. Energy proportionality? 100% 70% 12

13. 3#: Utilization < 10% … but most of the time here Already bad here 13

14. Sleeping saves energy Sleep-on-Idle (SoI) Devices enter sleep mode upon periods of inactivity 14

15. METRO SoI fails in access networks USER PARTISP PART ACCESS An ADSL line needs 1 minute to wake up … but cannot enjoy a minute’s sleep 15

16. What if we can put 80% of gateways to sleep? 15 W 100 W 1 W per modem Save big fraction at the user sideISPs … not so much 16

17. Line cards very unlikely to sleep by SoI Line cards Modem off DSLAM Modem on Static assignment of lines to DSLAM ports is a problem 17

18. ⟹ Greening the user part: aggregation ⟹ Greening the ISP part: line switching OUR APPROACH 18

19. Greening the user part – Aggregation On average 5-6 WiFi networks overlap in typical urban settings 19 Broadband Hitch-Hiking (BH 2 ) Threshold-based heuristic algorithm: direct traffic to neighbor gateways during light traffic conditions

20. Broadband Hitch-Hiking (BH 2 ) Load on home gateway is low  direct light traffic to a neighbor gateway and let home gateway sleep Load on neighbor gateway is low  look for another neighbor or go back to home gateway Load on neighbor gateway is high  go back to home gateway 20 15 W

21. BH 2 under the hood Gateways assumed to have: SoI Ability to wake on traffic BH 2 terminals use WiFi card virtualization: [Giustiniano et al., MobiCom ’10; Kandula et al., NSDI ’08] Estimate load on all gateways in range Maintain connectivity with 1 or few backup gateways Only modify the terminal wireless card driver 21

22. Greening the ISP part – Line switching DSLAM Line cards 22 AT&T operators connecting transatlantic calls at the international switchboard in New York, circa 1930. (www.corp.att.com) 40-way switch Full switching maximizes savings … but cost quickly grows with the number of ways Switching is for lines, not for packets

23. Small 4-way switches are enough Each k-switch packs active lines to the top DSLAM Line cards Put line cards to sleep Simple micro-electro-mechanical switches with near-zero power consumption 4-way switches 23

24. Evaluation Traces: CRAWDAD UCSD – 272 clients, 40 gateways Scenario: 4 x 12-port line cards; 5.6 avg. gateways Results for 4 schemes Baseline: no sleep – Gateways, modems, line cards don’t sleep – Each terminal only connects to its home gateway 24

25. How many gateways can sleep? Scheme 1: SoI for gateways, modems & line cards 25 Peak hours

26. How many gateways can sleep? BH 2 puts to sleep 70-90% of gateways Scheme 2: BH 2 + k-switch BH 2 every 150 s + 12 x 4-switches 26

27. How many gateways can sleep? Scheme 3: Optimal Best clients to gateways assignment + full switch 27

28. What is the impact of gateway density? Just home + 2 neighbors 50%+ of gateways sleep 28

29. How much energy can we save? BH 2 + k-switch saves 66% Optimal savings are 80% 29

30. What are the savings for the ISP? Scheme 4: SoI + k-switch Combine switching with SoI 30

31. What are the savings for the ISP? Savings come from both aggregation and switching 31

32. What are the savings for the ISP? Savings come from both aggregation and switching 32

33. Prototype deployment 1 st floor2 nd floor3 rd floor 33 Limitation: home gateway + 2 neighbor gateways per client

34. Incentives Security Privacy Considerations for deployment 34

35. A performance bonus Powering off lines makes the remaining … go faster due to reduced crosstalk! 35 Bonus: reduced crosstalk

36. Conclusions 80% energy can be saved at the access Aggregation + switching save 66% of energy Surprising result: turning DSL modems off increases the performance of remaining modems Applying our solution to all DSL users worldwide, yields savings of 33 TWh per year – ½ of energy going into US datacenters Thank you! Questions? – Or the output of 3 nuclear power plants 36