Real-Time Task Scheduling under Thermal Constraints

As the speed of integrated circuits increases, so does their power consumption. Most of this power is turned into heat, which must be dissipated effectively in order for the circuit to avoid thermal damage. Thermal control therefore has emerged as an important issue in design and management of circuits and systems. Dynamic speed scaling, where the input power is temporarily reduced by appropriately slowing down the circuit, is one of the major techniques to manage power so as to maintain safe temperature levels. In this study, we focus on thermally-constrained hard real-time systems, where timing guarantees must be met without exceeding safe temperature levels within the microprocessor. Speed scaling mechanisms provided in many of today?s processors provide opportunities to temporarily increase the processor speed beyond levels that would be safe over extended time periods. This dissertation addresses the problem of safely controlling the processor speed when scheduling mixed workloads with both hard-real-time periodic tasks and non-real-time, but latency-sensitive, aperiodic jobs. We first introduce the Transient Overclocking Server, which safely reduces the response time of aperiodic jobs in the presence of hard real-time periodic tasks and thermal constraints. We then propose a design-time (off-line) execution-budget allocation scheme for the application of the Transient Overclocking Server. We show that there is an optimal budget allocation which depends on the temporal character istics of the aperiodic workload. In order to provide a quantitative framework for the allocation of budget during system design, we present a queuing model and validate the model with results from a discrete-event simulator. Next, we describe an on-line thermally-aware transient overclocking method to reduce the response time of aperiodic jobs efficiently at run-time. We describe a modified Slack-Stealing algorithm to consider the thermal constraints of systems together with the deadline constraints of periodic tasks. With the thermal model and temperature data provided by embedded thermal sensors, we compute slack for aperiodic workload at run-time that satisfies both thermal and temporal constraints. We show that the proposed Thermally-Aware Slack-Stealing algorithm minimizes the response times of aperiodic jobs while guaranteeing both the thermal safety of the system and the schedulability of the real-time tasks. The two proposed speed control algorithms are examples of so-called proactive schemes, since they rely on a prediction of the thermal trajectory to control the temperature before safe levels are exceeded. In practice, the effectiveness of proactive speed control for the thermal management of a system relies on the accuracy of the thermal model that underlies the prediction of the effects of speed scaling and task execution on the temperature of the processor. Due to variances in the manufacturing of the circuit and of the environment it is to operate, an accurate thermal model can be gathered at deployment time only. The absence of power data makes a straightforward derivation of a model impossible. We, therefore, study and describe a methodology to infer efficiently the thermal model based on the monitoring of system temperatures and number of instructions used for task executions.