Yaoquan Zhong, Wei Guo, Weiqiang sun, Yaohui Jin and Weisheng Hu

1. Differentiated Protection Services with Failure Probability Guarantee for Workflow-Based Applications Yaoquan Zhong, Wei Guo, Weiqiang sun, Yaohui Jin and Weisheng Hu State Key Lab on Fiber-Optic Local Area Networks and Advanced Optical Communication Systems Shanghai Jiao Tong University, P.R.C E-mail: chris-zhong@sjtu.edu.cn

2. Outline • Motivation • Failure Probability • Differentiated Protection Services Algorithms • Simulation • Conclusion

3. Motivation • Failure Probability • Differentiated Protection Services Algorithms • Simulation • Conclusion

4. Optical Grid A set of distributed computing resources connected by WDM optical network • Applications: complex science and engineering computation (optical grid)

5. Workflow-based Application • An application consists of multiple dependent computation tasks and communication tasks node • Applicationis modeled as Directed Acyclic Graph (DAG) • Node: computation tasks • Edge: communication tasks edge

6. Execute an Application in Optical Grid Computation tasks  computing resources Communication tasks  lightpaths Workflow-based Application Optical Grid

7. Motivation • Failure Probability • Differentiated Protection Services Algorithms • Simulation • Conclusion

8. Failure probability • Failure Probability is the probability that optical link failure occurs in the execution of communication task and the workflow-based application. • There is a QoS requirement when user submits a workflow-based application. In this paper, Failure Probability is used. • The failure probability calculation of shared protected communication task is solved by employing conditional probability.

9. Single Lightpath Failure Probability Analysis The failure probability of single optical link: The failure probability of single lightpath:

10. Failure Probability in dedicated Protection Failure probability of the dedicated protected communication task: 10

11. Failure Probability in Shared Protection conditional probability: Note: dual-link failure is the key factor.

12. Motivation • Failure Probability • Differentiated Protection Services Algorithms • Simulation • Conclusion

13. Scheduling Objective • There is a tradeoff between high network resource utilization and low Failure Probability performance. • To minimize network resource cost while satisfying the Failure Probability requirement of application . Objective: the bandwidth the failure probability optical grid can provide to application transmission time the failure probability requirement from User lightpath length

14. Normal Differentiated Protection Services All of the communication tasks in one application are choose in SAME protection mode according to the Failure Probability requirement. Failure Probability requirement Cost-effective enough for workflow-based application？

15. MFP-DPS and ME-DPS Algorithms Maximal Failure Probability Differentiated Protection Services (MFP-DPS) algorithm FIRST chooses the communication task with max failure probability to improve its protection level. Maximal Efficiency Differentiated Protection Services (ME-DPS) algorithm FIRST chooses the communication task with max benefit-cost function to improve its protection level. Only some of the communication tasks in one application are choose to improve the protection level according to the Failure Probability requirement. more flexible and efficient in network resource allocation

16. Motivation • Failure Probability • Differentiated Protection Services Algorithms • Simulation • Conclusion

17. Simulation Environment Use the 26 nodes typical US network topology as our network infrastructure. • Failure probability requirements of applications are uniformly distributed over five classes {class 1, class 2, class 3, class 4 and class 5}, which respectively represent the failure probability {0.01, 0.001, 0.0001, 0.00001 and 0.000001}. • Randomly generate 100 applications (DAGs) in every failure probability requirement class. Failure repair time follows a negative exponential distribution with a mean value of 24 hours.

18. Resource overbuild Resource overbuild, which is defined as the amount of wavelength consumed by backup paths over the amount of wavelength utilized by working paths. Lower resource overbuild is preferred because it indicates lower resource consumption overhead and better backup-sharing optimization.

19. Average Finish Time Lower resource overbuild mean more network resource can be used to execute other communication tasks, and it makes the average finish time of the application shortened.

20. Protection Modes Distribution of Applications The applications distributionof NDPS algorithm Three different protection modes The proportions of shared-protected and dedicated-protected applications are increasing.

21. Protection Modes Distribution of Applications MFP-DPS and ME-DPS algorithms provide flexible protection modes to the workflow-based applications These two algorithms save the backup network resource through reducing the protection level of some communication tasks in an application when the failure probability requirement is satisfied. The applications distributionof MFP-DPS and ME-DPS algorithms

22. Motivation • Failure Probability • Differentiated Protection Services Algorithms • Simulation • Conclusion

23. Conclusion • Failure probability is a more accurate metric to measure security level of communication task in workflow-based application. • To ensure the security requirement, some communication tasks in applications may be not necessarily need any protection. • Simulation results show that ME-DPS algorithms is more efficient with lower resource overbuild and less finish time.

24. Thank you！