Energy Minimization For Pipelined Multicore Systems With Both Throughput And Latency Guarantees

Nowadays, embedded system is more and more powerful, the performance is greatly improved, the energy consumption of the system is more and more high, the problem of the system’s heat dissipation is more and more important. energy consumption has become the main bottleneck for the development of the embedded system. Thus, the energy-efficient scheduling for embedded real-time systems has become an important research in the embedded computing field.Scheduling streaming applications into a parallel pipeline on multicore architectures has become increasingly effective and widespread for classes of prevalent systems such as media and graphics processing, network packet processing systems. Because streaming applications are usually compute-intensive, they are energy hungry, which will cause problems if they are running in energy-constrained systems, such as battery-operated embedded devices. Thus, there is a great need to optimize energy consumption for streaming applications.This thesis study the energy efficient parallel-pipeline scheduling problem of streaming applications on multicores based on per-core DVS policy. A streaming application is modeled as a task graph, which is a weighted Direct Acyclic Graph (DAG). This thesis assume the application is scheduled into a parallel pipeline topology from the DAG based on a static scheduling policy. Existing studies with DVS technology are mostly use continuous frequency adjustment model, but this model does not apply to the actual system. Therefore, this thesis studies the discrete frequency adjustment model. The goal is to determine the optimal frequency assignment to minimize the energy consumption of streaming applications while guaranteeing two prime quality-of-service (QoS) requirements, i.e., throughput and latency.Firstly, this thesis study the energy minimization for pipelined multicore systems with both throughput and latency guarantees and proposes two different frequency assignment algorithms(PPS and PPC) while the processor model supports the restricted-scaling scheduling. PPS algorithm set the initial frequency of all tasks to the highest value, then gradually find out the task which reduces the most energy consumption In unit time and reduce its frequency, stretch its execution time till all tasks can’t stretch. PPC algorithm sets the initial frequency of all tasks to the minimum value, then gradually increase all tasks’s frequencys, compress the tasks’execution time until all tasks meet the throughput constraints, and then gradually find out the task which increases the least energy consumption In unit time and increase its frequency, shorten its execution time until all tasks’execution time meet the delay constraints. Then this thesis consider another processor model which supports the arbitrary-scaling scheduling. In this situation, this thesis first segment the parallel pipeline model reasonably, then base on the segment parallel pipeline model, this thesis propose another two frequency assignment algorithms:PPSS and PPSC. The simulation results show that the algorithms proposed in this paper could achieve remarkable results in energy saving and the effect for PPSS and PPSC algorithms are better than PPS and PPC algorithms.