Distributed sleep mode handling and task processing in massive multi-core processors

Abstract

This thesis addresses the problem of energy efficient real-time task processing in massive multi-core environments. A sleep mode handling method for processor architectures with support for per-core power gating is developed and the solution is inspired by cellular automata. In a cellular automaton each subcomponent has behavior based on local information which together makes up the behavior of the system. In order to evaluate the method, a simulator is tool is developed and used. Presented and evaluated are a set of qualitative measures for benchmarking the sleep mode handling policies. A key result is that it is possible to construct a distributed core sleep mode handling policy which only depends on local information e.g. the neighboring cores power states, and still performs comparable to policies that are not distributed. Simulation results show that it is possible to save a considerable amount of energy without extending the lateness compared to a system without power handling. The developed simulator tool is useful for understanding how the system behaves as the number of cores increases beyond 1000 and simulation results show that sudden changes in task load can affect lateness momentarily.