FairCloud: Sharing the Network in Cloud Computing

Computer Communication Review(2012) Arthur : Lucian Popa

Arvind Krishnamurthy Sylvia Ratnasamy Ion Stoica

Presenter : 段雲鵬

2/20

Outline

• Introduction• challenges sharing networks• Properties for network sharing• Mechanism• Conclusion

3/20

Some concepts

• Bisection bandwidth– Each node has a unit weight– Each link has a unit weight

• Flow def– standard five-tuple in packet headers

• B denotes bandwidth• T denotes traffic• W denotes the weight of a VM

4/20

Background

• Resource in cloud computing– Network , CPU , memory

• Network allocation– More difficult• Source, destination and cross traffic

– Tradeoff • payment proportionality VS bandwidth guarantees

5/20

Introduction

• Network allocation– Unkown to users , bad predictability

• Fairness issues– Flows, source-destination pairs, or sources alone ,

destination alone• Difference with other resource– Interdependent Users– Interdependent Resources

6/20

Assumption

• From a per-VM viewpoint• Be agnostic to VM placement and routing

algorithms• In a single datacenter• Be largely orthogonal to work on network

topologies to improve bisection bandwidth

7/20

Traditional Mechanism

• Per flow fairness– Unfair when simply instantiating more flow

• Per source-destination pair– Unfair when one VM communicates with more VMs

• Per source – Unfair to destinations

• Asymmetric – Only be fair for source or destination only

8/20

Examples

• Per source-destination pair

Per source

If there is little traffic on the A-F and B-E , B(A)=B(B) =B(E) =B(F) =2*B(C) =2*B(D) =B(G) =B(H)

B(E) =B(F) =0.25*B(D) , In the opposite direction, B(A) =B(B) =0.25*B(C)

9/20

Properties for network sharing(1)

• Strategy proofness– Can’t increase bandwidth by modifying behavior

at application level• Pareto Efficiency– X and Y is bottlenecked , when B(X-Y) increases,

B(A-B) must decrease ,otherwise congestion will be worse

A X Y B

A

1 M10 M 10 M

10/20

Properties for network sharing(2)

• Non-zero Flow Allocation– A strictly +B() between each pairs are expected

• Independence– When T2 increase , B1 should not be affected

• Symmetry– If all flows’ direction are swiched, the allocation should be the same

e.g

L2

L1

11/20

Network weight and user’s payment.

• Weight Fidelity(provide incentive)– Strict Monotonicity (Monotonicity)• If W(VM) increases ,then all its traffic must increase

(not decrease) .– Proportionality

• Guaranteed Bandwidth– Admission control

• They are conflicting, tradeoff

Subset P(2/3)

Subset Q(1/3)

No communication between P and Q

12/20

Per Endpoint Sharing (PES)

• Can explicitly trade between weight fidelity and guaranteed bandwidth

– NA denote the number of VMs A is communicating with

– WS-D=f(WS,WD) , WA-B=WB-A

– Normalized by L1 normalization• Drawback : Static Method (out of discussion)

13/20

Example

• WA-D=WA/NA+WD/ND =1/2+1/2=1

• WA-C=WB-D=1/2+1/1=1.5• Total Weight=4(4 VMs)• So WA-D=1/4=0.25 WA-C=WB-D=1.5/4=0.325

14/20

Comparison

15/20

PES

• For one host , B (closer VMs) instead of ∝(remote VMs)

• Higher guarantees for the worst case

• WA−B = WB−A =α*WA/NA+ β*WB/ NB

– α and β can be designed to weight between bandwidth guarantees and weight fidelity

16/20

One Sided PES (OSPES)

• Designed for tree-based topology

• WA−B = WB−A =α*WA/NA+ β*WB/ NB

• When closer to A, α = 1 and β = 0 • When closer to B, α = 0 and β = 1

17/20

OSPES

• fair sharing for the traffic towards or from the tree root – Resource allocation are depended on the root

– Non-strict monotonicity

When W(A) = W(B) , If the accesslink is 1 Gbs, then each VM is guaranteed 500 Mbps

WA-VM1=1/1 WB-VMi=1/10(i=2,3……,11)

18/20

Max-Min Fairness

• The minimum data rate that a dataflow achieves is maximized– The bottleneck is fully utilized

• Can be applied

19/20

Conclusion

• Problem : sharing the network within a cloud computing datacenter

• Tradeoff between payment proportionality and bandwidth guarantees

• A mechanism to make tradeoff between conflicting requirements

Reference• [1] Amazon web services.

http://aws.amazon.com.• [2] M. Al-Fares, A. Loukissas, and A. Vahdat. A

scalable, commodity data center• network architecture. In SIGCOMM. ACM,

2008.• [3] M. Al-Fares, S. Radhakrishnan, B. Raghavan,

N. Huang, and A. Vahdat.• Hedera: Dynamic Flow Scheduling for Data

Center Networks. In NSDI, 2010.• [4] H. Ballani, P. Costa, T. Karagiannis, and A.

Rowstron. Towards Predictable• Datacenter Networks. In ACM SIGCOMM,

2011.• [5] D. P. Bertsekas and R. Gallager. Data

networks (2. ed.). Prentice Hall, 1992.• [6] B. Briscoe. Flow rate fairness: Dismantling a

religion. ACM SIGCOMM

• Computer Communication Review, 2007.• [7] N. G. Duffield, P. Goyal, A. G.

Greenberg, P. P. Mishra, K. K. Ramakrishnan,

• and J. E. van der Merwe. A flexible model for resource management in virtual

• private networks. In SIGCOMM, 1999.• [8] A. Ghodsi, M. Zaharia, B. Hindman, et

al. Dominant resource fairness: fair• allocation of multiple resource types. In

USENIX NSDI, 2011.• [9] A. Greenberg, J. R. Hamilton, N. Jain,

S. Kandula, C. Kim, P. Lahiri, D. A.• Maltz, P. Patel, and S. Sengupta. VL2: A

Scalable and Flexible Data Center

21/20

Thanks !!