Towards Self-adaptive MPSoC Systems with Adaptivity Throttling

Today's multi-processor system-on-chip (MPSoC) systems increasingly have to deal with dynamically changing application workload scenarios. To cope with such dynamic application behavior, these systems could dynamically adapt the mapping of application tasks onto the underlying system resources to improve the system's performance. However, such performance improvement comes at the cost of a system reconfiguration in which application tasks may have to be migrated between processors. This trade-off implies that reconfiguring the system is only beneficial when the performance gains outweight the re-configuration overhead. To address this problem for MPSoCs, this paper presents a scenario-based run-time resource management framework with the ability of adaptivity throttling that uses the history of application scenario execution behavior to predict the actual benefit of a system reconfiguration to allow for explicitly deciding (at runtime) whether or not to reconfigure. Experimental results reveal that our proposed approach substantially improves the system's efficiency as compared to MPSoCs that do not provide such intelligent reconfiguration control.