Heuristic Resource Allocation Algorithms for Maximizing Allowable Workload

1. Heuristic Resource Allocation Algorithms for Maximizing Allowable Workload in Dynamic, Distributed, Real-Time Systems Presenter: David Fleeman { fleeman@ohio.edu } D. Juedes, F. Drews, L. Welch and D. Fleeman Center for Intelligent, Distributed & Dependable Systems Ohio University Athens, OH WPDRTS 2004 April 26, 2004

2. Motivating Example • Tasks which have workload dependent execution times. • Example originated in the generic air defense system • The detect task is periodic and identifies threats by filtering and evaluating radar tracks. • The engage task is event-driven and fires a missile at the threat • The guide missile task is event-driven periodic and uses sensor data to track a specific threat, recalculates flight path, and issues guidance commands to the missile. • Familiar “shooter-to-target” requirement. WPDRTS, April 26, 2004

3. Motivating Example (continued) • All three tasks have resource and timing requirements that are environment-dependent. • The detect task depends on both the number of radar tracks to process and the number of tracks that are actually threats. • The engage task is activated by events which occur at rates that are determined by the external environment. • The guide missile task depends on the number of missiles in flight. WPDRTS, April 26, 2004

4. Motivating Example (continued) • Traditional WCET analysis causes poor utilization of resources whenever there are little or no threats to be analyzed. • We characterize the resource needs of these tasks by execution profile functions that compute the resource needs as a function of workload. • These functions are used in this work to choose allocations that allow the applications: • To better utilize the resources. • Allow real-time constraints to be met • Minimize the need for reallocations which create system overhead at the worst possible times. WPDRTS, April 26, 2004

5. The Maximum Allowable Workload Problem (RMS) • Allocation of n independent tasks to m processors. • Running times of each task t is given as function of the system workload w. • Problem: Find an allocation of tasks to processors and a setting of w such that this allocation is feasible for all workloads w’≤ w, such that w is maximized. WPDRTS, April 26, 2004

6. Known Analytical Results • If the running-times of all of the tasks are well-behaved, then a modified version of First First is guaranteed to be within 41% of optimal, asymptotically. • If less than 12% of the system utilization is used up by input independent tasks (i.e., constant time tasks), then First First is within 33% of optimal, asymptotically. WPDRTS, April 26, 2004

7. A Modified Version of First Fit by Oh and Baker Input: <T,P> and a workload w. Output: An allocation alloc:T  P and “Feasible” or “Not Feasible” for each job i do place job i on the first processor j such that all tasks already assigned to processor j and task i can meet their deadlines when running with workload w. if no such processor j exists, return “Not Feasible” Return “Feasible” and alloc. WPDRTS, April 26, 2004

8. Using FF to Approximate MAW-RMS Use binary search to find a workload w such that the algorithm given on the previous page return “Feasible,” but the same algorithm returns “Not Feasible” for workload w+1. Use theallocation returned by the last feasible result of FF. WPDRTS, April 26, 2004

9. Experimental Results • We considered n=20,30,40,…,100 tasks • 10 non-identical processors, each of which described by its speed factor ranging within [10,30] • Periods of tasks were choosen from [2500,5000] • Random polynomials for workload functions • Choosen from WPDRTS, April 26, 2004

10. Experimental Results: 100.000 Iterations for Simulated Annealing and Random Search WPDRTS, April 26, 2004