Task Migration in Mesh NoCs over Virtual Point-to-Point Connections

Processor allocation in todays many core MPSoCs is a challenging task, especially since the order and requirements of incoming applications are unknown during design stage. To improve network performance, balance the workload across processing cores, or mitigate the effect of hot processing elements in thermal management methodologies, task migration is a method which has attracted much attention in recent years. Runtime task migration was first proposed in multicomputer with load balancing as the major objective. However, specific NoC properties such as limited amount of communication buffers, more sensitivity to implementation complexity, and tight latency and power consumption constraints bring new challenges in using task migration mechanisms in NoCs. As a consequence, the efficiency and applicability of traditional migration mechanisms (developed for multicomputers) are under question. Due to the limited resource budget in NoC-based MPSoCs as well as tight performance constraints of running applications, in this paper, we propose an efficient methodology based on virtual point-to-point (VIP for short) connections. These dedicated VIP connections provide low-latency and low-power paths for heavy communication flows created by task migration mechanisms. Analyzing the results show that the proposed scheme reduces message latency by 13% and migration latency by 14%, while 10% power savings can be achieved compared to the previously proposed task migration strategy (known as Gathering-Rout-Scattering) for mesh multiprocessors.