MMPI-based Task Migration Mechanism Design And Implementation In MPSoC

The NoC-based MPSoC has the advantages of low-power, scalability,flexibility, etc. To dynamically manage the computation and communicationresource at runtime, and improve resource utilization, this project designs andimplements a low-cost task migration mechanism for NoC-based MPSoC platform.In addition, the task migration mechanism is applied to node fault tolerance.First, the project designs and implements a low-cost task migration mechanism.The task migration mechanism depends on the MPSoC message passing interface(MMPI), in which program is independent of task mapping. Task is remapped byupdating task mapping table. Then task state information in uC/OS-II including taskstack and task control block is transferred to the migration destination node, onwhich a new task is created and the task state is loaded into the newly created task.Therefore, the migrated task can restore execution on a different node. Themechanism does not need task code transfer as well as checkpoints. Besides, tokeep correct inter-task communication, strategies targeted to different situations aredesigned to avoid influence that task migration has on inter-task communication.Then, based on the task migration mechanism, this project applies taskmigration to node fault tolerance in MPSoC. Fault tolerance is achieved byreplacing fault nodes with redundant nodes. Virtual topology layer, in charge ofconverting virtual topology to physical topology, is added into the MMPI layers.The distributed task migration strategy is implemented that each slave node sendstasks state information to the redundant node at intervals. In case a slave node isfailure, the tasks running on it can restore on the redundant node. In this method,the task is migrated from the faulty node to other one.Finally, the task migration mechanism is evaluated. The experimental resultsshow that without task code transfer and checkpoint, this task migration mechanismhas low migration delay and can response to the migration request immediately.Furthermore, the distributed task migration strategy supporting fault tolerance isimplemented on the platform. The experimental results show that the interval timesending task state has great influence on the task execution time.