MMPI-based Task Migration Algorithm Design And Implementation In Mpsoc

With the rapid development and application of the NoC-based multi-coresystems, uneven workload distribution and some damaged nodes occur across thenetwork. As a result, these cases can severely affect the performance of the entiresystem, and may even cause the entire burned chip due to localized overheating ordoes not work due to the damaged nodes. For such problems, tasks on the nodesoverloaded or damaged need to be migrated partly or entirely to other proper onesfor the purpose of workload balance or fault tolerance. In the process of migrationhowever it is difficult to avoid the vast communication cost.This paper proposes a workload criterion and a task migration algorithm on it.The workload criterion aims at avoiding the migration thrashing and reducingruntime of task. The task migration algorithm not only has the feature of workloadcriterion but also obtains a low migration communication cost. Firstly, thealgorithm calculates hop area of every overloaded node to obtain Candidate Load-light Node Matrix (CLNM) and Hop Matrix (HM). Secondly, it weights the HM byWeight Matrix (WM). Finally, finds out the optimal paired combination with theminimal migration communication cost according to the weighted HM and CLNM.Subsequently, in order to make the multi-core system support fault tolerance,in the paper, we firstly modify the multi-core system simulation platform developedin our laboratory and the MMPI parallel programming model. Then combine thetask migration algorithm proposed and a task migration mechanism to achieve faulttolerance in multi-core systems. When node fails, the system will migrate task onthe failed one to the proper one calculated by the task migration algorithm so as toachieve the purpose of fault tolerance.Finally, experiments and evaluations are made about the proposed taskmigration algorithm and task migration strategy based on fault tolerance.Experimental results show that the proposed algorithm reduces migrationcommunication cost approximately by18%compared with DP algorithm. Moreover,it can get lower energy consumption with an average reduction of3.9%in ourexperiments with four networks and more even energy distribution. Besides, thepaper correctly achieves the task migration strategy and fault tolerance in multi-core platform and show the main cost is the transmission of task status information.