Presentation is loading. Please wait.

Presentation is loading. Please wait.

Network Sharing Issues Lecture 15 Aditya Akella. Is this the biggest problem in cloud resource allocation? Why? Why not? How does the problem differ wrt.

Similar presentations


Presentation on theme: "Network Sharing Issues Lecture 15 Aditya Akella. Is this the biggest problem in cloud resource allocation? Why? Why not? How does the problem differ wrt."— Presentation transcript:

1 Network Sharing Issues Lecture 15 Aditya Akella

2 Is this the biggest problem in cloud resource allocation? Why? Why not? How does the problem differ wrt allocating other resources? FairCloud: Sharing the Network in Cloud Computing, Sigcomm 2012 What are the assumptions? Drawbacks?

3 Motivation Network?

4 Context Networks are more difficult to share than other resources X

5 Context Several proposals that share network differently, e.g.: – proportional to # source VMs (Seawall [NSDI11]) – statically reserve bandwidth (Oktopus [Sigcomm12]) –…–… Provide specific types of sharing policies Characterize solution space and relate policies to each other?

6 FairCloud Paper 1.Framework for understanding network sharing in cloud computing – Goals, tradeoffs, properties 2.Solutions for sharing the network – Existing policies in this framework – New policies representing different points in the design space

7 Goals 1.Minimum Bandwidth Guarantees – Provides predictable performance – Example: file transfer finishes within time limit A1A1 A2A2 Time max = Size / B min B min

8 Goals 1.Minimum Bandwidth Guarantees 2.High Utilization – Do not leave useful resources unutilized – Requires both work-conservation and proper incentives A BBB Both tenants activeNon work-conservingWork-conserving

9 Goals 1.Minimum Bandwidth Guarantees 2.High Utilization 3.Network Proportionality – As with other services, network should be shared proportional to payment – Currently, tenants pay a flat rate per VM  network share should be proportional to #VMs (assuming identical VMs)

10 Goals 1.Minimum Bandwidth Guarantees 2.High Utilization 3.Network Proportionality – Example: A has 2 VMs, B has 3 VMs A1A1 A2A2 Bw A B1B1 B3B3 B2B2 Bw B Bw A = 2 3 When exact sharing is not possible use max-min

11 Goals 1.Minimum Bandwidth Guarantees 2.High Utilization 3.Network Proportionality Not all goals are achievable simultaneously!

12 Tradeoffs Min Guarantee High Utilization Network Proportionality Not all goals are achievable simultaneously!

13 Tradeoffs Min Guarantee Network Proportionality

14 Tradeoffs Bw B Bw A AB Access Link L Capacity C Bw B = 11/13 C Bw A = 2/13 C 10 VMs Network Proportionality Bw A ≈ C/N T  0 #VMs in the network Bw B = 1/2 C Bw A = 1/2 C Minimum Guarantee Min Guarantee Network Proportionality

15 Tradeoffs High Utilization Network Proportionality

16 Tradeoffs L B1B1 B3B3 B2B2 B4B4 A1A1 A3A3 A2A2 A4A4 High Utilization Network Proportionality

17 Tradeoffs L B1B1 B3B3 B2B2 B4B4 A1A1 A3A3 A2A2 A4A4 Bw B = 1/2 C Bw A = 1/2 C Network Proportionality High Utilization Network Proportionality

18 Tradeoffs L B1B1 B3B3 B2B2 B4B4 A1A1 A3A3 A2A2 A4A4 P High Utilization Network Proportionality Uncongested path

19 Tradeoffs L B1B1 B3B3 B2B2 B4B4 A1A1 A3A3 A2A2 A4A4 Bw B Bw A < Network Proportionality L L Tenants can be disincentivized to use free resources If A values A 1  A 2 or A 3  A 4 more than A 1  A 3 Bw A +Bw A = Bw B L L P P High Utilization Network Proportionality Uncongested path

20 Network Proportionality Tradeoffs L B1B1 B3B3 B2B2 B4B4 A1A1 A3A3 A2A2 A4A4 P Network proportionality applied only for flows traversing congested links shared by multiple tenants High Utilization Uncongested path Congestion Proportionality

21 Network Proportionality Tradeoffs L B1B1 B3B3 B2B2 B4B4 A1A1 A3A3 A2A2 A4A4 Uncongested path P Bw B Bw A = Congestion Proportionality L L High Utilization Congestion Proportionality

22 Network Proportionality Tradeoffs Still conflicts with high utilization High Utilization Congestion Proportionality

23 Tradeoffs B1B1 B3B3 A1A1 A3A3 B2B2 B4B4 A2A2 A4A4 L2L2 C 1 = C 2 = C Network Proportionality High Utilization L1L1 Congestion Proportionality

24 Tradeoffs B1B1 B3B3 A1A1 A3A3 B2B2 B4B4 A2A2 A4A4 L1L1 L2L2 Bw B Bw A = Congestion Proportionality L1 Bw B Bw A = L2 Network Proportionality High Utilization C 1 = C 2 = C Congestion Proportionality

25 Tradeoffs B1B1 B3B3 A1A1 A3A3 B2B2 B4B4 A2A2 A4A4 L1L1 L2L2 Demand drops to ε Network Proportionality High Utilization C 1 = C 2 = C Congestion Proportionality

26 Tradeoffs B1B1 B3B3 A1A1 A3A3 B2B2 B4B4 A2A2 A4A4 ε C - ε ε Tenants incentivized to not fully utilize resources Network Proportionality High Utilization C 1 = C 2 = C Congestion Proportionality L1L1 L2L2

27 Tradeoffs B1B1 B3B3 A1A1 A3A3 B2B2 B4B4 A2A2 A4A4 ε C - 2 ε Uncongested ε C - ε L1L1 L2L2 C 1 = C 2 = C Network Proportionality High Utilization Tenants incentivized to not fully utilize resources Congestion Proportionality

28 L2L2 Tradeoffs B1B1 B3B3 A1A1 A3A3 B2B2 B4B4 A2A2 A4A4 ε C - 2 ε C/2 Uncongested C 1 = C 2 = C Network Proportionality High Utilization Congestion Proportionality Tenants incentivized to not fully utilize resources L1L1

29 Tradeoffs Proportionality applied to each link independently Congestion Proportionality Link Proportionality Network Proportionality High Utilization L2L2 B1B1 B3B3 A1A1 A3A3 B2B2 B4B4 A2A2 A4A4 L1L1

30 L2L2 Tradeoffs B1B1 B3B3 A1A1 A3A3 B2B2 B4B4 A2A2 A4A4 Full incentives for high utilization Congestion Proportionality Link Proportionality Network Proportionality High Utilization L1L1

31 Goals and Tradeoffs Min Guarantee Link Proportionality High Utilization Congestion Proportionality Network Proportionality

32 Guiding Properties Min Guarantee Link Proportionality High Utilization Congestion Proportionality Network Proportionality Break down goals into lower-level necessary properties

33 Properties Min Guarantee Link Proportionality High Utilization Congestion Proportionality Network Proportionality

34 Work Conservation Bottleneck links are fully utilized Static reservations do not have this property

35 Properties Min Guarantee Link Proportionality High Utilization Congestion Proportionality Network Proportionality Work Conservation

36 Utilization Incentives Tenants are not incentivized to lie about demand to leave links underutilized Network and congestion proportionality do not have this property Allocating links independently provides this property

37 Properties Min Guarantee Link Proportionality High Utilization Congestion Proportionality Network Proportionality Work Conservation Utilization Incentives

38 Communication-pattern Independence Allocation does not depend on communication pattern Per flow allocation does not have this property – (per flow = give equal shares to each flow)  Same Bw 

39 Properties Min Guarantee Link Proportionality High Utilization Congestion Proportionality Network Proportionality Work Conservation Utilization Incentives Comm-Pattern Independence

40 Symmetry Swapping demand directions preserves allocation Per source allocation lacks this property – (per source = give equal shares to each source)  Same Bw 

41 Goals, Tradeoffs, Properties Min Guarantee Link Proportionality High Utilization Congestion Proportionality Network Proportionality Work Conservation Utilization Incentives Comm-Pattern Independence Symmetry

42 Outline 1.Framework for understanding network sharing in cloud computing – Goals, tradeoffs, properties 2.Solutions for sharing the network – Existing policies in this framework – New policies representing different points in the design space

43 Per Flow (e.g. today) Min Guarantee Link Proportionality High Utilization Congestion Proportionality Network Proportionality Work Conservation Utilization Incentives Comm-Pattern Independence Symmetry

44 Per Source (e.g., Seawall [NSDI’11]) Min Guarantee Link Proportionality High Utilization Congestion Proportionality Network Proportionality Work Conservation Utilization Incentives Comm-Pattern Independence Symmetry

45 Static Reservation (e.g., Oktopus [Sigcomm’11]) Min Guarantee Link Proportionality High Utilization Congestion Proportionality Network Proportionality Work Conservation Utilization Incentives Comm-Pattern Independence Symmetry

46 New Allocation Policies 3 new allocation policies that take different stands on tradeoffs

47 Proportional Sharing at Link-level (PS-L) Min Guarantee Link Proportionality High Utilization Congestion Proportionality Network Proportionality Work Conservation Utilization Incentives Comm-Pattern Independence Symmetry

48 Proportional Sharing at Link-level (PS-L) Per tenant WFQ where weight = # tenant’s VMs on link A B WQ A = #VMs A on L Bw B Bw A = #VMs A on L #VMs B on L Can easily be extended to use heterogeneous VMs (by using VM weights)

49 Proportional Sharing at Network-level (PS-N) Min Guarantee Link Proportionality High Utilization Congestion Proportionality Network Proportionality Work Conservation Utilization Incentives Comm-Pattern Independence Symmetry

50 Proportional Sharing at Network-level (PS-N) Congestion proportionality in severely restricted context Per source-destination WFQ, total tenant weight = # VMs

51 Proportional Sharing at Network-level (PS-N) WQ A1  A2 = 1/N A1 + 1/N A2 A1A1 A2A2 N A2 N A1 Total WQ A = #VMs A Congestion proportionality in severely restricted context Per source-destination WFQ, total tenant weight = # VMs

52 Proportional Sharing at Network-level (PS-N) WQ B WQ A = #VMs A #VMs B Congestion proportionality in severely restricted context Per source-destination WFQ, total tenant weight = # VMs

53 Proportional Sharing on Proximate Links (PS-P) Min Guarantee Link Proportionality High Utilization Congestion Proportionality Network Proportionality Work Conservation Utilization Incentives Comm-Pattern Independence Symmetry

54 Assumes a tree-based topology: traditional, fat-tree, VL2 (currently working on removing this assumption) Proportional Sharing on Proximate Links (PS-P)

55 Assumes a tree-based topology: traditional, fat-tree, VL2 (currently working on removing this assumption) Min guarantees – Hose model – Admission control Proportional Sharing on Proximate Links (PS-P) A1A1 Bw A1 A2A2 Bw A2 AnAn Bw An …

56 Assumes a tree-based topology: traditional, fat-tree, VL2 (currently working on removing this assumption) Min guarantees – Hose model – Admission control High Utilization – Per source fair sharing towards tree root Proportional Sharing on Proximate Links (PS-P)

57 Assumes a tree-based topology: traditional, fat-tree, VL2 (currently working on removing this assumption) Min guarantees – Hose model – Admission control High Utilization – Per source fair sharing towards tree root – Per destination fair sharing from tree root Proportional Sharing on Proximate Links (PS-P)

58 Deploying PS-L, PS-N and PS-P Full Switch Support – All allocations can use hardware queues (per tenant, per VM or per source-destination) Partial Switch Support – PS-N and PS-P can be deployed using CSFQ [Sigcomm’98] No Switch Support – PS-N can be deployed using only hypervisors – PS-P could be deployed using only hypervisors, we are currently working on it

59 Evaluation Small Testbed + Click Modular Router – 15 servers, 1Gbps links Simulation + Real Traces – 3200 nodes, flow level simulator, Facebook MapReduce traces

60 Many to one Bw B Bw A AB N One link, testbed PS-P offers guarantees Bw A N

61 MapReduce One link, testbed Bw B Bw A 5 R5 M M+R = 10 M Bw B (Mbps) PS-L offers link proportionality

62 MapReduce Network, simulation, Facebook trace

63 MapReduce Network, simulation, Facebook trace PS-N is close to network proportionality

64 MapReduce Network, simulation, Facebook trace PS-N and PS-P reduce shuffle time of small jobs by 10-15X

65 Summary Sharing cloud networks is not trivial First step towards a framework to analyze network sharing in cloud computing – Key goals (min guarantees, high utilization and proportionality), tradeoffs and properties New allocation policies, superset properties from past work – PS-L: link proportionality + high utilization – PS-N: restricted network proportional – PS-P: min guarantees + high utilization What are the assumptions, drawbacks?


Download ppt "Network Sharing Issues Lecture 15 Aditya Akella. Is this the biggest problem in cloud resource allocation? Why? Why not? How does the problem differ wrt."

Similar presentations


Ads by Google