Energy-aware scheduling in heterogeneous computing systems
In the last decade, the grid computing systems emerged as useful provider of the computing power required for solving complex problems. The classic formulation of the scheduling problem in heterogeneous computing systems is NP-hard, thus approximation techniques are required for solving real-world s...
Saved in:
| Main Author: | |
|---|---|
| Format: | masterThesis |
| Language: | English |
| Published: |
2013
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/20.500.12008/33370 |
| Tags: |
No Tags, Be the first to tag this record!
|
| Summary: | In the last decade, the grid computing systems emerged as useful provider of the computing power required for solving complex problems. The classic formulation of the scheduling problem in heterogeneous computing systems is NP-hard, thus approximation techniques are required for solving real-world scenarios of this problem. This thesis tackles the problem of scheduling tasks in a heterogeneous computing environment in reduced execution times, considering the schedule length and the total energy consumption as the optimization objectives. An efficient multithreading local search algorithm for solving the multi-objective scheduling problem in heterogeneous computing systems, named MEMLS, is presented. The proposed method follows a fully multi-objective approach, applying a Pareto-based dominance search that is executed in parallel by using several threads. The experimental analysis demonstrates that the new multithreading algorithm outperforms a set of fast and accurate two-phase deterministic heuristics based on the traditional MinMin. The new ME-MLS method is able to achieve significant improvements in both makespan and energy consumption objectives in reduced execution times for a large set of testbed instances, while exhibiting very good scalability. The ME-MLS was evaluated solving instances comprised of up to 2048 tasks and 64 machines. In order to scale the dimension of the problem instances even further and tackle large-sized problem instances, the Graphical Processing Unit (GPU) architecture is considered. This line of future work has been initially tackled with the gPALS: a hybrid CPU/GPU local search algorithm for efficiently tackling a single-objective heterogeneous computing scheduling problem. The gPALS shows very promising results, being able to tackle instances of up to 32768 tasks and 1024 machines in reasonable execution times. |
|---|