SORT

Abstract

This article proposes a Semi Online Reliable Task (SORT) mapping approach to many-core platforms divided into two sections: offline and online. The offline section is a twofolded approach. It maintains the reliability of the mapped task graph against soft errors considering the reliability threshold defined by designers. As wear-out mechanisms decrease the lifetime of the system, our proposed approach increases the lifetime of the system using task migration scenarios. It specifies task migration plans with the minimum overhead using a novel heuristic approach. SORT maintains the required level of reliability of the task graph in the whole lifetime of the system using a replication technique with minimum replica overhead, maximum achievable performance, and minimum temperature increase. The online segment uses migration plans obtained in the offline segment to increase the lifetime and also permanently maintains the reliability threshold for the task graph during runtime. Results show that the effectiveness of SORT improves on bigger mesh sizes and higher reliability thresholds. Simulation results obtained from real benchmarks show that the proposed approach decreases design-time calculation up to 4,371% compared to exhaustive exploration while achieving a lifetime negligibly lower than the exhaustive solution (up to 5.83%).