Mecanismo de controle de QoS através de DFS em MPSOCS

Abstract

The quality of service (QoS) management in NoC-based MPSoCs, with dozens of applications executing simultaneously, is an open research challenge in the integrated circuit design area. Adaptability techniques, which use different QoS metrics, have been used at design time to guarantee the QoS of the applications. Designers include in their systems monitoring schema that guides embedded controllers in managing the resources of the MPSoC to satisfy the QoS requirements imposed to the applications. In order words, MPSoCs are able to self-adapt while running a set of applications. The self-adaptation capability is a fundamental characteristic to satisfy the QoS requirements on the systems with dynamic workload. The dynamic voltage and frequency scaling (DVFS) is the most used adaptation method for reducing the overall energy consumption of an MPSoC. However, this method does not take into account other QoS requirements such as throughput or latency. Another example of adaptation technique is task migration, whose main goal is to balance the workload of the MPSoC. The QoS control mechanism proposed in the scope of this Thesis uses the dynamic frequency scaling (DFS) technique to control the QoS parameters of the application, keeping energy consumption low profile. Each processor has a monitoring system, a QoS evaluation system and an adaptation module, which are used to control the QoS parameters to satisfy the QoS requirements imposed to the applications. At the system startup, each processor uses a DFS policy that tries to optimize the communication with its neighbor s processors. The processors use this policy up to the moment when they reach a steady frequency state. After reaching the steady frequency state the QoS monitoring starts, evaluating if they the requirements imposed at design time are respected. The proposed QoS control mechanism was evaluated using two synthetic and one real application, using the HeMPS MPSoC, with the throughput and latency parameters as the QoS parameters to be controlled. The presented results show that the proposed QoS control mechanism can satisfy the imposed QoS requirements using the DFS technique while maintaining low energy consumption on the HeMPS MPSoC.