Research On Real-Time Task Scheduling Approaches For Energy Consumption And Temperature Optimization

With the pursuing of high performance computing and the continuous development of chip manufacturing technology,more and more functional elements are integrated on the same board,resulting in higher power density.Meanwhile,the severe thermal issues caused by the high power density restrict the further improvement of performance in.turn.High power consumption reduces the life time of the system,while high temperature significantly affects the reliability and safety of the system.Besides,there exists a vicious circle relationship between power consumption and temperature:the higher the power consumption,the higher the temperature;the higher the temperature,the higher the leakage power consumption,i.e.,the higher the power consumption.Further,due to the limited equipment space,both the energy capacity and cooling technique of the real-time systems are restricted;energy consumption and temperature optimization by hardware design is limited.To this end,it is necessary to explore the optimization approach to minimize both the higher energy consumption and temperature caused by the continuously released real-time tasks while considering the strong dependency between temperature and power consumption.For a given computing system,the hardware-dependent powerand thermal features are fixed,and the task scheduling is the most intuitive and efficient method to optimize both energy consumption and temperature.Task scheduling depends on the specific computing system and the tasks to be scheduled.High performance computing systems are usually multi-processor systems,which can be homogeneous or heterogeneous.Tasks can be independent or constrained if there exist execution-dependent relationships between tasks.To this end,under the background of real-time system development,this dissertation studies the energy-temperature optimized real-time task scheduling approaches for independent tasks and constrained tasks(including mixed-criticality tasks and precedence-constrained tasks)on homogeneous systems and heterogeneous systems.The main contents together with their innovations are shown as follows.Firstly,we study the scheduling problem for independent periodic tasks on homogeneous real-time systems,and proposes an energy-temperature optimal GPSS scheduling approach for real-time tasks.For this problem,existing works only analyze the temperature profile under steady state,and also ignore the effect of task executing sequence on temperature.To this end,this work first deduces the energy-temperature optimal condition for Global Processor Sharing Scheduling(GPSS)which holds in both steady state and transient state based on the proposed end-temperature-equivalent task construction method,and generalizes it to other scheduling algorithms.Then,proposes a thermal-aware method to determine the task execution sequence based on the thermal law exhibited by the power hetero-geneity of tasks.Finally,based on the above optimization,proposes an energy-temperature optimal GPSS scheduling approach that satisfies the processor speed constraint and task timing constraint.Extensive experiments validate the efficiency of the proposed task construction method,energy-temperature optimal condition and the GPSS scheduling approach based on the optimal condition.Secondly,we study the scheduling problem for Mixed-Criticality tasks on homogeneous real-time systems,and proposes an energy-temperature optimal mixed-criticality fluid scheduling approach for real-time tasks.Mixed-Criticality System(MCS)is a new emerging real-time system,which integrates applications with different criticality levels on the same platform so as to reduce the hardware cost.Existing related works most focus on improving the scheduling efficiency rather than the energy optimization.A few energy saving works also ignore the temperature-dependent leakage power.To this end,this work first analyzes the schedulability conditions under different execution modes of MCS,then deduces the energy-temperature optimal speed assignments under different execution modes,and finally demonstrates through thorough experiments that the proposed task scheduling approach can better minimize both the energy consumption and temperature while guaranteeing acceptable schedulability ratio.Thirdly,we study the scheduling problem for independent periodic tasks on heterogeneous real-time systems,and proposes an energy/thermal balanced two-phase scheduling approach for real-time tasks,which minimizes the system energy consumption and temperature by narrowing the energy-temperature gap between different processors and different tasks on the same processor.In existing related works,the task assignment approaches usually only take the dynamic power consumption as the metric,with temperature as the constraint;the task scheduling approaches most only conduct optimization at the processor-level speed scaling.Both of them cannot fully exploit the optimization space of energy consumption and temperature.To this end,considering the heterogeneity of both processors and tasks,this work first proposes an energy-temperature aware task assignment heuristic to balance the energy/thermal loads of processors,and then leveraging the task-level speed scaling technique to balance the energy/thermal loads of tasks on the same processor.Extensive experiments demonstrate that the proposed two-phase task scheduling approach can better minimize both the energy consumption and temperature of the heterogeneous real-time systems compared with the state-of-the-art approaches.Finally,we study the task scheduling problem for precedence-constrained application on heterogeneous real-time systems,and proposes an energy-temperature optimized scheduling approach for real-time systems.This approach minimizes the system energy consumption and temperature by balancing the energy/thermal loads of processors,balancing the optimization space of tasks,and by reducing the waiting time between different tasks.Existing works usually ignore the temperature-dependent leakage power,only focusing on the minimization of dynamic energy consumption.However,with the increasing proportion of leakage power,any optimization approaches ignoring this part should be inefficient in the deep sub-micro era.To this end,this work first analyzes the task scheduling method for the tasks with precedence constraints(usually denoted by a directed acyclic graph(DAG)),deduces the metric to evaluate the energy-temperature optimization efficiency of processors,and based on which proposes a task assignment heuristic by balancing the energy/thermal load of processors and the optimization space of tasks.Then,for the given task assignment,further minimize the energy consumption and temperature of the processor by reducing the waiting time between tasks having the same successor task.Experiments conducted on real-life benchmarks demonstrate that the proposed task scheduling approach can better minimize both the energy consumption and temperature compared with the state-of-the-art approaches.To sum up,this dissertation studies the energy-temperature optimized real-time task scheduling problems for independent/constrained tasks on homogeneous/heterogeneous systems.By taking the strong leakage/temperature dependency into consideration,this work proposes task scheduling approaches to fully explore the optimization space for both the energy consumption and temperature of real-time systems.Both theoretical analysis and experimental results demonstrate the efficiency and superiority of the proposed real-time task scheduling approaches.The research results can be widely used in various real-time systems,such as the mobile computing systems,flight management systems,intelligent transportation systems and environment monitoring systems.In our future work,we will extend this research in much wider scenarios,like the more than two-criticality systems and multiple DAG applications.