36
An Autonomic Framework in Cloud Environment Jiedan Zhu Advisor: Prof. Gagan Agrawal

An Autonomic Framework in Cloud Environment

  • Upload
    zizi

  • View
    31

  • Download
    0

Embed Size (px)

DESCRIPTION

An Autonomic Framework in Cloud Environment. Jiedan Zhu Advisor: Prof. Gagan Agrawal. Outline. Introduction Motivation Application Framework Design Overview Key Components Experiments Conclusion. Outline. Introduction Motivation Application Framework Design Overview Key Components - PowerPoint PPT Presentation

Citation preview

Page 1: An Autonomic Framework in Cloud Environment

An Autonomic Framework in Cloud Environment

Jiedan Zhu

Advisor: Prof. Gagan Agrawal

Page 2: An Autonomic Framework in Cloud Environment

Outline

• Introduction• Motivation Application• Framework Design– Overview– Key Components

• Experiments• Conclusion

Page 3: An Autonomic Framework in Cloud Environment

Outline

• Introduction• Motivation Application• Framework Design– Overview– Key Components

• Experiments• Conclusion

Page 4: An Autonomic Framework in Cloud Environment

Introduction

• Cloud Computing• various computation and storage resources• pay-as-you-go

• User Constraints• Execution Time• Cost

• Problems• under utilization most of the time• longer execution• more expensive than as expected

Page 5: An Autonomic Framework in Cloud Environment

Main Challenges

• Possible Solution• server consolidation ------ task consolidation• live migration ------ light-weighted migration

• Our Work• an autonomic framework

– Three techniques for three kinds of prior knowledge

• Our Goals• Keep the application to complete within the time

constraint• Keep the cost within the cost budget

Page 6: An Autonomic Framework in Cloud Environment

Contributions

• Our Contributions– our system can save the cost up to 59% and more

cost-efficient compared to the case when there is no resource scheduling

– effective and adaptive on different workflow structures

– performs better with the prior knowledge of CPU and memory requirements of tasks

Page 7: An Autonomic Framework in Cloud Environment

Outline

• Introduction• Motivation Application• Framework Design– Overview– Key Components

• Experiments• Related Work• Conclusion

Page 8: An Autonomic Framework in Cloud Environment

Motivation Application

• Volume Rendering– DAG-based Workflow– Parallelism

• Constraints– machine capacities– resource contention– varying time constraint

& cost budget

Page 9: An Autonomic Framework in Cloud Environment

• Figures and level

Motivation Application

Page 10: An Autonomic Framework in Cloud Environment

Outline

• Introduction• Motivation Application• Framework Design– Overview– Key Components

• Experiments• Related Work• Conclusion

Page 11: An Autonomic Framework in Cloud Environment

Framework Overview

Component 2

Component 3 Component 4

Component 1

Page 12: An Autonomic Framework in Cloud Environment

Key Components 1

• Task Monitoring Agent– task status information• CPU usage, memory usage, iteration #, iteration time

– checkpoints for each task• paths of input and output data• parameters for workflow• intermediate states such as iteration variable

Page 13: An Autonomic Framework in Cloud Environment

Key Components 2

• Progress Analysis Module– analyze the execution progress– workflow-specific prior knowledge

A: CPU and memory requirements of tasks» initial assignment plan

B: iteration structures of the workflowC: iteration structures of the tasks

Page 14: An Autonomic Framework in Cloud Environment

Progress Estimation -- A

• CPU and Memory Requirements of Tasks• wocExecTime, wocTaskTime• pastTime, e.g. 500 sec

estTaskTimet

estLevelTimet

estFutureTimei+1n

pastTime

current level is 2

e.g. reqCPU: 50%, curCPU: 20%, so the ratiot is 2.5e.g. ratiot is 2.5, wocTaskTimet is 300 sec, estTimet is 750 sec, only 1 task on current level 2, so estLevelTimei is 750 sec

future level 3: task 4 wocTaskTime is 100sec, task5 wocTaskTime is 10 sec, task 6 is 300 sec, so estLevelTime3 is 100 x 2.5 = 250 sec, estLevelTime4 is 300 x 2.5 = 750 sec

Total is 500 + 750 + 250 + 750 = 2250 sec

Page 15: An Autonomic Framework in Cloud Environment

• Iteration Structures of The Workflow• the jth iteration, total iterations is k• wocLevelTimei

• pastLevelTime1i e.g. 500 sec

Progress Estimation -- B

estLevelTimel

pastLevelTime1i

estFutureTimei+1n

e.g. total is 3 iterations, now it is the 1st iteration, pastLevelTime is for both level 1 and level 2. reqLevelTime1 is 150 sec and reqLevelTime2 is 250 sec. so ratio1

i is 500 / 400 = 1.25

current level is 3

e.g. current level is level 3 and reqLevelTime3 is 300 sec, so estLevelTime3 is 375 sec

The time for the 1st iteration is 500 + 375 + 312.5 = 1187.5 sec, so total is 3562.5 sec

future level is level 4, reqLevelTime4 is 250 sec, so estLevelTime4 is 312.5 sec

Page 16: An Autonomic Framework in Cloud Environment

Progress Estimation -- C

• Iteration Structures of Tasks• wocLevelTimei suppose no iteration structures of workflow

• remainIterNumt, avgTPerItert,pastTime

e.g. 500, 0.02 sec, 500 sec

estFutureTimei+1n

estComTime1i

estLevelTimel

current level 2

e.g. the remaining execution time for task 3 is 500 x 0.02 = 10 sec. Only 1 task on level 2, so the completion time for both level 1 and 2 is 500 + 10 = 510 sec. reqLevelTime1 and reqLevelTime2 are 150 sec and 250 sec, so ratio1

i is 1.275

future level 3: reqLevelTime3is 100sec, and for level 4, estLevelTime4 is 300 sec, so estLevelTime3 is 100 x 1.275 = 127.5 sec, estLevelTime4 is 300 x 1.275 = 382.5 sec, so estFutureTime3

4 is 510 sec

Total is 510 + 510 = 1020 sec.

Page 17: An Autonomic Framework in Cloud Environment

Progress Estimation

Page 18: An Autonomic Framework in Cloud Environment

Key Components 3

• Scheduling Module– Greedy Algorithm• if the time constraint can not be satisfied

– reschedule the instances

• if the cost budget can not be satisfied while the time constraint is satisfied– consolidate the tasks and reduce the number of instances

vm1 vm2 New vm

Page 19: An Autonomic Framework in Cloud Environment

Key Components

• Migration Module• light-weighted checkpoints ------ migration overhead is

small• timing for migration ------ 10 second point• keep data dependencies and resume the

communication ------ global address book

Page 20: An Autonomic Framework in Cloud Environment

Outline

• Introduction• Motivation Application• Framework Design– Overview– Key Components

• Experiments• Related Work• Conclusion

Page 21: An Autonomic Framework in Cloud Environment

Experiment Design

• Experiment Goals– system effectiveness evaluation– system performance comparison under different workflow-

specific prior knowledge

• Experiment Environment– instance type ------ c1.medium

• 2 virtual cores• 1.7GB memory• Moderate I/O performance

– pricing• $0.17 / hour ------ $0.17 / 10 seconds

Page 22: An Autonomic Framework in Cloud Environment

Experiment Design

• Real Application – Volume Rendering

• Synthetic Workflows– synthetic workflow 1– synthetic workflow 2– synthetic workflow 3

Page 23: An Autonomic Framework in Cloud Environment

Experiment Design

• Synthetic workflow 1– Number of parallelism

is static– No iteration structures

of workflow

Page 24: An Autonomic Framework in Cloud Environment

Experiment Design

• Synthetic workflow 2– the number of parallelism is varying– no iteration structures of the workflow

Page 25: An Autonomic Framework in Cloud Environment

Experiment Design

• Synthetic workflow 3– Iteration structures for both workflow and tasks

Page 26: An Autonomic Framework in Cloud Environment

Experiment 1

• System Effectiveness Evaluation– our system vs. no scheduling– on synthetic workflow 1 and 2

Page 27: An Autonomic Framework in Cloud Environment

Experiment 1

Page 28: An Autonomic Framework in Cloud Environment

Experiment Results

• Experiment Conclusion– our system can save up to 59% cost and more

cost-efficient compared to the case when there is no resource scheduling

– effective ------ satisfying all user requirements– adaptive to workflows of different structures

Page 29: An Autonomic Framework in Cloud Environment

Experiment 2

• system performance comparison under different workflow-specific prior knowledge

vrCM vrIter performance-price ratio comparisons

Page 30: An Autonomic Framework in Cloud Environment

Experiment 2

syn1CM syn1Iter performance-price ratio comparisons

syn2CM syn2Iter performance-price ratio comparisons

Page 31: An Autonomic Framework in Cloud Environment

Experiment 2

syn3CM syn3Iter performance-price ratio comparisons

Page 32: An Autonomic Framework in Cloud Environment

Experiment Results

• Experiment Conclusion– With the prior knowledge with CPU and memory

requirements of task, our system performs better in terms of smoothness and performance-price ratio than with other prior knowledge• may benefit from initial assignment plan

Page 33: An Autonomic Framework in Cloud Environment

Outline

• Introduction• Motivation Application• Framework Design– Overview– Key Components

• Experiments• Related Work• Conclusion

Page 34: An Autonomic Framework in Cloud Environment

Related Work• Amazon web services. http://aws.amazon.com/.• Y. Ajiro and A. Tanaka. Improving packing algorithms for server consolidation. In

CMG-CONFERENCE-, volume 2, page 399. Computer Measurement Group; 1997, 2007.

• L. Chen, Q. Zhu, and G. Agrawal. Supporting dynamic migration in tightly coupled grid applications. In SC 2006 Conference, Proceedings of the ACM/IEEE, pages 28–28. IEEE, 2006.

• Q. Zhu and G. Agrawal. Resource provisioning with budget constraints for adaptive applications in cloud environments. In Proceedings of the 19th ACM International Symposium on High Performance Distributed Computing, pages 304–307. ACM, 2010.

• Q. Zhu, J. Zhu, and G. Agrawal. Power-aware consolidation of scientific workflows in virtualized environments. In Proceedings of the 2010 ACM/IEEE International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1–12. IEEE Computer Society, 2010.

Page 35: An Autonomic Framework in Cloud Environment

Outline

• Introduction• Motivation Application• Framework Design– Overview– Key Components

• Experiments• Related Work• Conclusion

Page 36: An Autonomic Framework in Cloud Environment

Conclusion

• Autonomic framework in the Cloud Environment

• Three techniques for three kinds of prior knowledge

• Task consolidation and light-weighted migration

• Effective, adaptive and save the cost up to 59%