1. Application-level Resource Provisioning
Gurmeet Singh, Carl Kesselman, Ewa Deelman

2. Outline Motivation Model Comparison of 2 provisioning algorithms
Current work

3. Motivation Resource Provisioning
Reduces uncertainty associated with workflow performance
Allows deterministic performance
Enables user based workload scheduling
Allows co-allocation of resources
Improves performance as compared to scheduling alone for certain class of applications.

4. Outline Motivation Model Comparison of 2 provisioning algorithms
Current work

5. System Model
Ascertaining the resource availability is central to the provisioning model.
r = 1,…,R Grid sites in the system.
Query Model: Er(n,d) = {<s1,c1,f1>,…,<sk,ck,fk>}
Publish Model: Er(.) = {<s1,n1,d1,c1,f1>,…,<sk,nk,dk,ck,fk>}
Global Resource Availability

6. Slots with FCFS scheduling

7. Slots with backfilling

8. Provisioning Model Allocation Plan : set of resource slots
a = { rs1, rs2, …., rsn } С A
Workflow G = (V,E) to be scheduled.
The Workflow can consist of a scheduled and an unscheduled part.
V = Vs U Vu
Both the AP and the schedule can be build incrementally

9. Allocation and Scheduling
Available Slots
Workflow
Scheduled Workflow over allocated slots.

10. Goal
The goal is to identify and allocate an allocation plan, a, in order to optimize the following costs
Total allocation cost
Total scheduling cost
Scheduling cost = SCω(a) = makespan of workflow G over a using ω.
A single cost metric is created using a weighted sum of the two costs.

11. Outline Motivation Model Comparison of 2 provisioning algorithms
Current work

12. Min-Min provisioning algorithm
Modify the level-based Min-Min algorithm to minimize the total cost at each step instead of only the makespan.
Also, do the allocation along with the scheduling.

13. Cost Metric Performance

14. GA Use a genetic algorithm based search
A solution is a n (= |A|) bit binary string.
Simple encoding and decoding.
Selection
Crossover
Mutation
Elitism
Search using a fixed population size and certain number of iterations.
GA paired with a simple list scheduling algorithm.

15. % reduction in total cost

16. Findings GA does a better overall allocation optimization than Min-Min
Min-Min outperforms GA for low value of α.
Min-Min performs worse with high values of α due to its non backtracking nature.
GA performance depends on the cardinality of the global set A.
The cost metric performs considerable allocation optimization for a little loss in makespan.
“Application level Resource Provisioning” 2nd IEEE Intl Conf on E-Science and Grid Computing, Dec 2006.

17. Outline Motivation Model Comparison of 2 provisioning algorithms
Current work

18. Current Work
Comparison of provisioned and non-provisioned approach with a trace driven simulation.
Multi-Objective GA (MOGA)
Slot Generation mechanisms for real Grid schedulers.