Sharing Cloud Networks

Transcription

1 Sharing Cloud Networks Lucian Popa, Gautam Kumar, Mosharaf Chowdhury Arvind Krishnamurthy, Sylvia Ratnasamy, Ion Stoica UC Berkeley

2 State of the Cloud Network??? 2

3 Guess the Share TCP Per flow Per Source Per Destination Per- VM Proportional 3 : 1 1 : 1 2 : 1 3 : 2 A 1 A 2 A 3 B 1 Link L B 2 Alice : Bob =? :? 3

4 Challenges Network share of a virtual machine (VM) V depends on» Collocated VMs,» Placement of destination VMs, and» Cross- traffic on each link used by V Network differs from CPU or RAM» Distributed resource» Usage attribution (source, destination, or both?) Traditional link sharing concepts needs rethinking 4

5 Requirements Min Bandwidth Guarantee Aggregate Proportionality High Utilization Introduce performance predictability Network shares proportional to the number of VMs Do not leave bandwidth unused if there is demand 5

6 Requirement 1: Guaranteed Minimum B/W Provides a minimum b/w guarantee for each VM Captures the desire of tenants to get performance isolation for their applications B mina B minb B minx A B X All VMs 6

7 Requirement 2: Aggregate Proportionality Shares network resources across tenants in proportion to the number of their VMs Captures payment- proportionality» Similar to other resources like CPU, RAM etc. Desirable properties» Strategy- proofness: Allocations cannot be gamed» Symmetry: Reversing directions of flows does not change allocation 7

8 Design Space Min Bandwidth Guarantee Aggregate Proportionality High Utilization Symmetry Strategy- Proofness 8

9 Requirement 3: High Utilization Provides incentives such that throughput is only constrained by the network capacity» Not by the inefficiency of the allocation or by disincentivizing users to send traffic Desirable properties» Work Conservation: Full utilization of bottleneck links» Independence: Independent allocation of one VM s traffic across independent paths 9

10 Design Space Tradeoff 1 Min Bandwidth Guarantee Aggregate Proportionality High Utilization Symmetry Strategy- Proofness Independence Work Conservation 10

11 Tradeoff 1: Min B/W vs. Proportionality A1 B1 C/2 C/3 2C/3 C/2 Link L with Capacity C A2 B2 B3 Share of Tenant A can decrease arbitrarily! 11

12 Design Space Tradeoff 1 Tradeoff 2 Min Bandwidth Guarantee Aggregate Proportionality High Utilization Symmetry Strategy- Proofness Independence Work Conservation 12

13 Tradeoff 2: Proportionality vs. Utilization A 1 A 2 C/4 A 3 C/4 A 4 B 1 B 3 C/4 C/4 Link L with Capacity C B 2 B 4 To maintain proportionality, equal amount of traffic must be moved from A1- A2 to A1- A3 => Underutilization of A1- A3 13

14 Per- link Proportionality Restrict to congested links only Share of a tenant on a congested link is proportional to the number of its VMs sending traffic on that link 14

15 Per Endpoint Sharing (PES) Five identical VMs (with unit weights) sharing a Link L A 1 A 2 A 3 B 1 Link L B 2 15

16 Per Endpoint Sharing (PES) Resulting weights of the three flows: N A = 2 A 1 B 1 3/2 3/2 2 Link L A 2 A 3 B 2 W A N A W B N B To generalize, weight of a flow A- B on link L is W A- B =

17 Per Endpoint Sharing (PES) W Symmetric A W W A- B = + B = W N A N B- A B Proportional» sum of weights of flows of a tenant on a link L = sum of weights of its VMs communicating on that link Work Conserving Independent Strategy- proof on congested links 17

18 Generalized PES Scale weight of A by α A L B Scale weight of B by β W A W A- B = W B- A = α NA + β W B N B α > β if L is more important to A than to B (e.g., A s access link) 18

19 One- Sided PES (OSPES) Scale weight of A by α A L B Scale weight of B by β W A W A- B = W B- A = α NA + β W B N B α = 1, β = 0 if A is closer to L α = 0, β = 1 if B is closer to L Highest B/W Guarantee Per Source closer to source Per Destination closer to destination *In the Hose Model 19

20 Comparison Per Flow Per Source Static Reservation Link PES OSPES Link Proportionality K K J J K Symmetry Strategy- Proofness Utilization J J K J J Independence Work Conservation B/W Guarantee L L J L K 20

21 Full- bisection B/W Network Aggregate Bandwidth PerFlow PerSource Link PES Network OSPES PES Tenant1 Tenant2 Ideal Tenant 1 has one- to- one communication pattern Tenant 2 has all- to- all communication pattern 21

22 MapReduce Workload Aggregate Bandwidth Tenant 1 Tenant 2 Tenant 3 Tenant 4 Tenant 5 Per Flow Per Source Link PES Network OSPES PES Optimal W1:W2:W3:W4:W5 = 1:2:3:4:5 22

23 Summary Sharing cloud networks is all about making tradeoffs» Min b/w guarantee VS Proportionality» Proportionality VS Utilization Desired solution is not obvious» Depends on several conflicting requirements and properties» Influenced by the end goal 23