Achieving QoS in NoC-based MPSoCs through Dynamic Frequency Scaling

Abstract

The management of Quality-of-Service (QoS) constraints in NoC-based MPSoCs, with dozens of tasks running simultaneously, is still a challenge. Techniques applied at design or run-time to address this issue adopts different QoS metrics. Designers include in their systems monitoring techniques, adapting at run-time the QoS parameters to cope with the required constraints. In order words, MPSoC are able to self-adapt themselves, while executing a given set of applications. Self-adaptation capability is a key feature to meet applications' requirements in dynamic systems. Dynamic Voltage and Frequency Scaling (DVFS) is an adaptation technique frequently used to reduce the overall energy consumption, not coupled to QoS constraints, as throughput or latency. Another example of adaptation technique is task migration, which focus on throughput or latency optimization. The self-adaptation technique proposed in this paper adopts Dynamic Frequency Scaling (DFS) trading-off power consumption and QoS constraints. Each processor running the applications' tasks initially reaches a steady state leading each task to a frequency level that optimizes the communication with neighbor tasks. The goal of the initial state is to reach a trade-off between power consumption and communication throughput. Next, the application performance is monitored to adjust the frequency level of each task according to the QoS parameters. Results show that the proposed self-adaptability scheme can meet the required QoS constraints, by changing the frequency of the PEs running the application tasks.