1. Job scheduling algorithm based on Berger model in cloud environment Advances in Engineering Software (2011) Baomin Xu,Chunyan Zhao,Enzhao Hua,Bin Hu 2013/1/251 Shang-Chi Wu

2. Outline Introduction Berger model Job scheduling algorithm based on Berger model Experiment Conclusions and questions 2013/1/25 2

3. Cloud computing It is a new pattern User-oriented design which provides varied services to meet the needs of different users 2013/1/25 3

4. Cloud computing The basic mechanism of cloud computing is to dispatch the computing tasks to resource pooling A variety of applications gain computing power, storage and a variety of software services according to their needs 2013/1/25 4

5. Cloud computing 2013/1/25 5 Task

6. Cloud computing The commercialization and the virtualization technology adopted by cloud computing has poured into new features for cloud architecture. Cloud computing needs pay more attention to the fairness of resources allocation 2013/1/25 6

7. Relation The relations between resource supply and demand are similarities with commodity economy model The resources provider->commodity supplier The resources users ->commodity buyers 2013/1/25 7

8. Relation The resources allocation in cloud computing has a very good fit with the distribution theory of social wealth 2013/1/25 8

9. Berger model The Berger model of distributive justice is based on expectation states Expectation states –Actor how to generate expectations of itself and other individuals –Expectations how to affect the behavior of actor 2013/1/25 9

10. Berger model Characteristic C is any aspect of a person A goal-object, GO, is any object that an actor might want, namely expectation 2013/1/25 10

11. Berger model Definition of distributive justice evaluation function in Berger model AR (Actual Reward) –actor obtains actually JR (Just Reward) –actor comparison of the justice distribution of other ordinary person 2013/1/25 11

12. Job scheduling algorithm based on Berger model Map the theory of distributive justice in Berger model to resource allocation model in cloud computing 2013/1/25 12

13. Task classification based on QoS In cloud computing, QoS is a metrics of user satisfaction with cloud services –Completion time: For real-time demand higher users –Bandwidth: User requests need higher communication bandwidth 2013/1/25 13

14. Fairness constraint C x user tasks E x expectation resources go x actual allocation resources C x is the QoS characteristics GO X is the general expectation 2013/1/25 14

15. Justice Task Justice –The justice evaluation function of task System Justice –Task set:T={T 1,T 2,…, T n } –Justice function set:J={J 1,J 2,…,J n } 2013/1/25 15

16. Description of tasks and resources Resource characteristics set of the virtual machine VM i C i = { C i1, C i2, C ir }, r = 1, 2, 3 The performance vector of Vm i VM i = [ EC i1, EC i2, EC i3 ] 2013/1/25 16 CPU memory bandwidth

17. The general expectations function The general expectations vector of the type e i = [ e i1, e i2, e i3 ] When the task has multiple expectations preference –mathematical expectation 2013/1/25 17 weight for CPU weight for memory weight for bandwidth

18. Completion time For Task T i Total number of virtual machine VM = { VM 1, …, VM m } Produce a candidate collection VM i = { VM 1, …, VM t }, t<m Select a virtual machine from Vm i according to the general expectation 2013/1/25 18

19. Completion time general expectations 2013/1/25 19

20. The normalized value 2013/1/25 20

21. Euclidean distance The minimum distance means the similarity of the two is best Vector X = [X 1, …, X n ] Vector Y = [Y 1, …, Y n ] 2013/1/25 21

22. Pseduo code 2013/1/25 22

23. the general expectations function 2013/1/25 23 Justice evaluation function When |J i | > 1, the system automatically adjusts the vector of general expectations

24. Bandwidth For Task T i Total number of virtual machine VM = { VM 1, …, VM m } Produce a candidate collection VM i = { VM 1, …, VM t }, t<m Select a virtual machine from Vm i according to the general expectation 2013/1/25 24

25. When |J i | > 1, the system automatically adjusts the vector of general expectations Justice evaluation function 2013/1/25 25 Justice evaluation function

26. Algorithm description 2013/1/25 26

27. The first type: e 1 = [0.7, 0.1, 0.2] The second type: e 2 = [0.3, 0.2, 0.5] Algorithm 1: job scheduling algorithm based on Berger model Algorithm 2: job scheduling algorithm based on the optimal completion time Experiment 2013/1/25 27

28. Experimental data 2013/1/25 28

29. Execution efficiency of algorithm 1’s slightly worse than algorithm 2’s Algorithm 1’s completion time is better than algorithm 2’s in task 0-3 Analysis of experimental results 2013/1/25 29

30. Algorithm 1 is able to better meet user expectations Analysis of experimental results 2013/1/25 30

31. Algorithm 1 enables task 0–3 to obtain good computing power with better fairness Analysis of experimental results 2013/1/25 31

32. Algorithm 1 can better meet the task preferences with better fairness Analysis of experimental results 2013/1/25 32

33. Conclusions and questions 2013/1/25 33 Berger model theory on distributive justice in the field of social distribution was first introduced into the job scheduling The proposed algorithm in this paper is effective implementation of user tasks, and with better fairness

34. Conclusions and questions 2013/1/25 34 The initial value of general expectation vector is empirical value Build a fuzzy neural network