Energy-aware scheduling of multi-version tasks on heterogeneous real-time systems
| Autor: | Julius Roeder, Clemens Grelck, Sebastian Altmeyer, Benjamin Rouxel |
|---|---|
| Přispěvatelé: | Parallel Computing Systems (IvI, FNWI) |
| Rok vydání: | 2021 |
| Předmět: |
Heuristic (computer science)
Computer science Real-time computing 020207 software engineering 02 engineering and technology Energy consumption Scheduling (computing) Task (computing) 020204 information systems 0202 electrical engineering electronic engineering information engineering ddc:004 Frequency scaling Integer programming Energy (signal processing) Efficient energy use |
| Zdroj: | SAC The 36th Annual ACM Symposium on Applied Computing: Virtual Conference : March 22–March 26, 2021, 501-510 STARTPAGE=501;ENDPAGE=510;TITLE=The 36th Annual ACM Symposium on Applied Computing Proceedings of the 36th Annual ACM Symposium on Applied Computing |
| DOI: | 10.1145/3412841.3441930 |
| Popis: | The emergence of battery-powered devices has led to an increase of interest in the energy consumption of computing devices. For embedded systems, dispatching the workload on different computing units enables the optimisation of the overall energy consumption on high-performance heterogeneous platforms. However, to use the full power of heterogeneity, architecture specific binary blocks are required, each with different energy/time trade-offs. Finding a scheduling strategy that minimises the energy consumption, while guaranteeing timing constraints creates new challenges. These challenges can only be met by using the full heterogeneous capacity of the platform (e.g. heterogeneous CPU, GPU, DVFS, dynamic frequency changes from within an application). We propose an off-line scheduling algorithm for dependent multiversion tasks based on Forward List Scheduling to minimise the overall energy consumption. Our heuristic accounts for Dynamic Voltage and Frequency Scaling (DVFS) and enables applications to dynamically adapt voltage and frequency during run time. We demonstrate the benefits of multi-version task models coupled with an energy-aware scheduler. We observe that selecting the most energy efficient version for each task does not lead to the lowest energy consumption for the whole application. Then we show that our approach produces schedules that are on average 45.6% more energy efficient than schedules produced by a state-of-the-art scheduling algorithm. Next we compare our heuristic against an optimal solution derived by an Integer Linear Programming (ILP) formulation (deviation of 1.6% on average). Lastly, we empirically show that the energy consumption predicted by our scheduler is close to the actual measured energy consumption on a Odroid-XU4 board (at most-15.8%). |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|