Efficient Run-time Resource Management of a Manycore Accelerator for Stream-based Applications

As the computation complexity of user applications is steadily increasing, it becomes popular to equip a many-core accelerator in a mobile embedded system. And the system status may change dynamically due to various factors such as workload variation, QoS requirement change, and unexpected component failure. In this paper, we address the problem of how to manage the dynamic behavior of the applications on a manycore-based accelerator for energy minimization. To this end, we propose a novel resource management technique that guarantees the throughput requirement of stream-based multimedia applications assuming that each operation mode of a multimedia application is specified by a synchronous dataflow. In the proposed approach, at design-time, we determine throughput-maximized mappings of each SDF graph by varying the number of allocated processors. Then, at run-time, we exploit the pre-computed mapping information to adjust the mapping of active applications to the processors without user intervention on the system status change. We validate the proposed scheme through intensive experiments with an in-house simulator built on top of Noxim, a Network-on-Chip simulator. Experimental results show the enhanced adaptability to dynamic system status change compared to other state-of-the-art approaches. We also quantitatively evaluate the overheads of the proposed run-time management technique in terms of computation, energy, and communication traffic involved.