A surface tension and coalescence model for dynamic distributed resources allocation in Massively Parallel Processors on-Chip

Massively Parallel Processors on-Chip, presenting the same problems of their non-monolithic counterparts, exacerbated by the limited on-chip resources, are the most challenging architectures in the processor architectures domain. In this paper, a novel nature-inspired decentralized algorithm, aiming at the definition of clusters of processors to be assigned to different threads, is presented and evaluated. Taking inspiration from liquid surface tension and drops coalescence, the proposed solution achieves better performances than other distributed solutions, reducing fragmentation and communication latency within the clusters.