Greening Backbone Networks

30-minute presentation for the Green IT and Smart Grid course at Stony Brook University, on "Greening Backbone Networks" by Fisher et. al published at the first SIGCOMM green networking workshop in 2010.

Transcript

1. Greening Backbone Networks Emaad Ahmed Manzoor CSE691: Green IT & Smart Grid March 2016 Will Fisher, Martin Suchara, Jennifer Rexford First Workshop on Green Networking, SIGCOMM ’10
2. Backbone Networks
3. https://en.wikipedia.org/wiki/File:Internet_map_1024.jpg
4. https://commons.wikimedia.org/wiki/File:NSFNET-traffic-visualization-1991.jpg
5. Aggregate Links (2-40 per-cable)
6. http://www.klalaboratories.com/service/asset/network_cabling_3_lg.jpg
7. https://www.reddit.com/r/mildlyinteresting/comments/3w14gh/this_giant_cable_stuffed_with_hundreds_of_wires/
8. Router Line Cards
9. https://farm4.staticflickr.com/3572/3567497578_aa942f35dd_b.jpg
10. Not power-proportional https://farm4.staticflickr.com/3572/3567497578_aa942f35dd_b.jpg
11. Power-proportional Router: Architectural Design and Experimental Evaluation, Bin Liu et. al. IWQoS 2014. Not power-proportional
12. Cannot be put to “sleep”* Not power-proportional https://farm4.staticflickr.com/3572/3567497578_aa942f35dd_b.jpg
13. Cannot be put to “sleep”* Not power-proportional *(2010) https://farm4.staticflickr.com/3572/3567497578_aa942f35dd_b.jpg
14. Average link-utilisation in backbone networks is 30-40%.
15. Can we power-down links (line cards) and adjust data flow based on traffic demand?
16. Problem formalisation Naive solution Integer linear program Efficient heuristics Evaluation Future
17. Problem formalisation Naive solution Integer linear program Efficient heuristics Evaluation Future
18. Input
19. G(V, E) s t w u v c(u,t) c(w,v) c(v,t) c(s,w) c(s,u) Topology Input
20. G(V, E) s t w u v c(u,t) c(w,v) c(v,t) c(s,w) c(s,u) Topology D (s1, t1, h2) (s2, t2, h2) . . . Demands Input
21. G(V, E) s t w u v c(u,t) c(w,v) c(v,t) c(s,w) c(s,u) Topology D (s1, t1, h2) (s2, t2, h2) . . . Demands Cables per link B Input
22. Output
23. s t w u v c(u,t) c(w,v) c(v,t) c(s,w) c(s,u) Flow of demand d on each link fd(s,u) fd(s,w) fd(w,v) fd(v,t) fd(u,t) fd(u, v) Output
24. s t w u v c(u,t) c(w,v) c(v,t) c(s,w) c(s,u) Flow of demand d on each link fd(s,u) fd(s,w) fd(w,v) fd(v,t) fd(u,t) Total flow on link (u,v) f(u, v) = X d fd(u, v) fd(u, v) Output
25. Problem formalisation Naive solution Integer linear program Efficient heuristics Evaluation Future
26. min number of powered cables s.t. link loads ≤ capacities flow conservation carries all traffic demands
27. min s.t. link loads ≤ capacities flow conservation carries all traffic demands X (u,v)2E f(u, v)
28. min s.t. flow conservation carries all traffic demands X (u,v)2E f(u, v) X D fd(u, v)  c(u, v), 8(u, v) 2 E
29. min s.t. carries all traffic demands X (u,v)2E f(u, v) X D fd(u, v)  c(u, v), 8(u, v) 2 E X v2V fd(u, v) = X w2V fd(w, u), 8d, 8u 6= sd, td
30. min s.t. X (u,v)2E f(u, v) X D fd(u, v)  c(u, v), 8(u, v) 2 E X v2V fd(sd, v) = X w2V fd(w, td) = hd, 8d X v2V fd(u, v) = X w2V fd(w, u), 8d, 8u 6= sd, td
31. min number of powered cables s.t. link loads ≤ capacities flow conservation carries all traffic demands Models what if links are energy proportional? • Fractional linear program • Bounds on power savings
32. Upper bound on energy savings f(u,v) = 6.3 Gbps (solution of FLP) c(u,v) = 10 Gbps (given topology) B = 10 (given cables per link) — Round down: 6 Gbps — Turn off: 4 cables
33. Lower bound on energy savings f(u,v) = 6.3 Gbps (solution of FLP) c(u,v) = 10 Gbps (given topology) B = 10 (given cables per link) — Round up: 7 Gbps — Turn off: 3 cables
34. Unfortunately, the lower bound solution is feasible but can be extremely suboptimal.
35. (S, Ti, ✏) (Bi, Bi+1, ✏)
36. Problem formalisation Naive solution Integer linear program Efficient heuristics Evaluation Future
37. The bundle size B needs to be explicitly taken into account.
38. min X (u,v)2E nuv Minimise the sum of the number of active cables on each link f(u, v)  nuv B c(u, v), 8(u, v) 2 E Number of active cables need to be sufficient for the demand
39. min X (u,v)2E nuv Minimise the sum of the number of active cables on each link f(u, v)  nuv B c(u, v), 8(u, v) 2 E Number of active cables need to be sufficient for the demand NP-Complete — D2CIF
40. Problem formalisation Naive solution Integer linear program Efficient heuristics Evaluation Future
41. Fast Greedy Heuristic 1. Solve FLP - Obtain f(u,v), round up. 2. Find (u,v) with the most unused capacity - Remove one cable from the link. - Solve the FLP with the new capacities. Feasible? Remove a cable from the link (u,v).
42. Fast Greedy Heuristic 1. Solve FLP - Obtain f(u,v), round up. 2. Find (u,v) with the most unused capacity - Remove one cable from the link. - Solve the FLP with the new capacities. Feasible? Remove a cable from the link (u,v). Not feasible? Mark link (u,v) as final.
43. Fast Greedy Heuristic 1. Solve FLP - Obtain f(u,v), round up. 2. Find (u,v) with the most unused capacity - Remove one cable from the link. - Solve the FLP with the new capacities. Feasible? Remove a cable from the link (u,v). Not feasible? Mark link (u,v) as final. Repeat from 2, ignoring all final links
44. Exhaustive Greedy Heuristic Remove the cable that - Retains solution feasibility. - Minimises the marginal penalty of each removal. Marginal penalty - Increases as excess traffic is rerouted along longer paths. - Measures as the FLP objective difference after cable removal.
45. Bi-level Greedy Heuristic Exhaustive greedy heuristic with pairs of cables.
46. Problem formalisation Naive solution Integer linear program Efficient heuristics Evaluation Future
47. FGH EGH BGH
48. Abilene Topology
52. 54% savings with B = 1 73% savings with B = 3
54. Problem formalisation Naive solution Integer linear program Efficient heuristics Evaluation Future
55. https://tools.ietf.org/html/draft-zhang-greennet-01 Impact
56. Dynamic demands Real-world traffic traces Approximation algorithms (?) Routing constraints
57. .