A User Experience-based Cloud Service Redeployment Mechanism KANG Yu Yu Kang, Yangfan Zhou, Zibin...

Preview:

Citation preview

A User Experience-based Cloud Service Redeployment

Mechanism

KANG Yu

Yu Kang, Yangfan Zhou, Zibin Zheng, and Michael R. Lyu{ykang,yfzhou,

zbzheng,lyu}@cse.cuhk.edu.hk

Department of Computer Science & Engineering

The Chinese University of Hong KongHong Kong, China

School of Computer ScienceNational University of Defence Technology

Changsha, China

Introduction

Overview of Cloud-based Services

Redeploying Service Instances

Experiment

Obtaining User Experience

Conclusion and Future Work

2CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Introduction

Cloud Computing Systems–Auto scaling

Dynamic allocation of computing resources

–Elastic load balanceDistributes and balances the incoming traffic

3CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Introduction

• Typical approach of auto scaling and load balance (Amazon EC2)

4CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Introduction

Current approaches are not optimized for users–Auto scaling

Do not consider distributions of the end users

–Elastic load balance Do not take the user specifics (e.g.,

user location) into considerations

5CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Introduction

• Our contribution:–User experience model in cloud –A new service redeployment method

• Two advantages:1)Improve auto scaling techniques

Launch best set of service instances

2)Extend elastic load balance Directs user request to a nearby one.

6CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Introduction

Overview of Cloud-based Services

Redeploying Service Instances

Experiment

Obtaining User Experience

Conclusion and Future Work

7CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Framework of Cloud-Based Services

• Data centers• Instances• Users

8CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Framework of Cloud-Based Services

• Round Trip Time (RTT) can be kept by the cloud provider.

• User experience contains three elements:1. Internet delay between a user and a

cloud data center (This is the most significant part)

2. Delay inside the data center3. Time to process the service request

9CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Framework of Cloud-Based Services

10CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Challenges of Hosting the Cloud Services

• Difficult to foresee user experience

• Delay can be measured (should take advantage of it)

11CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Introduction

Overview of Cloud-based Services

Redeploying Service Instances

Experiment

Obtaining User Experience

Conclusion and Future Work

12CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Obtaining User Experience

• Measuring Internet delay–RTT can be recorded

• Predict the Internet Delay–Not every data center is visited–Find similar users and predict the

connection.

13CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Obtaining User Experience

14CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Introduction

Overview of Cloud-based Services

Redeploying Service Instances

Experiment

Obtaining User Experience

Conclusion and Future Work

15CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Minimize Average Cost

16

Given:Z = the set of data centersC = the set of usersdij = distance between every pair (i,j) ∈ C╳Z

Minimize:

Subject to:𝑍′ ⊂ 𝑍, ∣𝑍′∣ = 𝑘

N

1i'

min ijZjd

CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Minimize Average Cost

17CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Minimize Average Cost

• k-median problem • NP-hard• W[2]-hard with k as parameter• W[1]-hard with capacity l as

parameter• In FPT with both as parameter

algorithm: O(f(k,l)no(1)) time

18CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Minimize Average Cost

• Approximate Algorithms:1. Exhaustive Search2. Greedy Algorithm3. Local Search Algorithm (3 + ε

approximation)4. Random Algorithm

19CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Problems of the Model

• Local Optimizer• Number of users connected to an

instance• Acceptable whenever response time

less than a threshold T

20CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Maximize Close User Amount

21

Given Bipartite graph 𝐵(𝑉1,��2,𝐸) where

∣𝑉1∣ = 𝑀, ∣𝑉2∣ = 𝑁 ∈ 𝑖 𝑉1, ∈ 𝑗 𝑉2

(𝑖, 𝑗) ∈ 𝐸, 𝑑𝑖𝑗 ≤ 𝑇;

(𝑖, 𝑗) ∉ 𝐸, otherwise.Maximize:

∣𝑁𝐵(𝑉′)∣Subject to:

𝑉 ′ ⊂ 𝑉1, ∣ 𝑉 ′∣ = 𝑘CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Maximize Close User Amount

22

{v1,v2,v3,v5}

v1 v2 v3 v4 v5

{v1,v2,v4}{v1,v3,v4}

{v4,v5}

CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Maximize Close User Amount

• Max k-cover problem• NP-hard• W[2]-hard with k as parameter• W[2]-hard (general) and FPT (tree-

like) with maximum subset size as parameter

• FPT if both maximum subset size and capacity as parameter

23CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Maximize Close User Amount

• Approximate Algorithms:1. Greedy Algorithm (1-1/e

approximation)2. Local Search Algorithm

24CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Introduction

Overview of Cloud-based Services

Redeploying Service Instances

Experiment

Obtaining User Experience

Conclusion and Future Work

25CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Dataset Description

• Deploy our WSEvaluator to 303 distributed computers of PlanetLab invoke to 4302 the Internet services

• A 303 * 4302 matrix containing response-time values

26CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Introduction

Overview of Cloud-based Services

Redeploying Service Instances

Experiment

Obtaining User Experience

Conclusion and Future Work

27CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Conclusion and Future Work

• Our work–A framework of new features –Formulate the redeployment problems.

• Future Work–Formulate the network capability in

detail–Optimize initial service instances

deployment

28CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Q & A

29CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Necessity of Redeployment

30CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Weakness of Auto Scaling

31CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Comparing Algorithms for k-Median

32CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Comparing Algorithms for k-Median

• Theoretical time complexity– Exhaustive search:– Greedy:– Local Search:

)( NMO k

)( NMkO

)( NMkO tt 33CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Redeployment Algorithms for Max k-Cover

• 20 instances are selected to provide service for 4000 users.

• Expect 200 per server.

34CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Redeployment Algorithms for Max k-Cover

• compare the average cost: max k-cover v.s. k-median

35CLOUD 2011, Washington DC, USA, July 4 - 9, 2011

Recommended