| With the extension of embedded system application area, higher requirements are put forward to the performance of embedded systems by application requirements, which act as the core power of embedded systems development. Equipped with high performance and high integration, multi-processor architecture has been provided and widely used. Advance in semiconductor technology has impelled multicore processor chips to be more high-integrated. However, when the feature size decreases to a certain range, the number of processors remaining on the chip increases the power consumption density within the chip, and heat production per unit area has increased drastically, which results in high chip temperature. The fact of overheating affects system performance and reliability seriously, which also increased the cost of cooling system. Therefore, the on-chip thermal effect has been extensively concerned and studied in recent years because of its importance and urgency. Among all the researches, the research on task mapping and power management methods to meet the temperature-related constraints is one of the hot issues. Based on the studies of basic dynamic thermal management technologies and methods towards many-core systems, this paper proposed a thermal-aware dynamic task mapping algorithm and real-time power management method.Firstly, a real-time thermal-ware task mapping algorithm TWeNA is proposed based on CoNA, which sets the system temperature and performance as dual-optimization target. Among the algorithm design, a reasonable task sorting algorithm is proposed based on communications weights of the task graph to optimize communicating performance on the network in design phase. Moreover, based on the real-time temperature feedback, and taking the smart hill climbing first node selection algorithm as base, the criterion of first-node selecting is optimized which infects the mapping area choice. During the process of task mapping, a new analysis method towards priorities of the candidate nodes is presented, taking comprehensive consideration of the impact on system temperature and performance.During the operation process of system, the distribution of workloads and thermal changes variably. This paper proposed a reasonable dynamic power management scheme under the constraint of TDP(Thermal Design Power), based on the monitoring information of the system and regular temperature feedback. The technology of DVFS(dynamic voltage frequency scaling) is employed to make the run-time scheme of fine-grained power management, which aims at optimizing system performance and temperature. On the phase of scheme design, a scheduling priority analysis procedure of the on-going applications and system performance impact factor are proposed. The overall scheduling process of power management is presented. Through analyzing the requirements and functions of the power control unit, the overall architecture design of the thermal-aware power management on many-core system is presented.On the stage of simulation, a multi-core system verification platform is built to evaluate the proposed task mapping algorithms and power management mechanism. The results showed that, TWeNA algorithm improved the network throughput by about 20% compared with the state-of-the-art task mapping algorithm, while the network packet-delivering delay has also been reduced. On the aspect of temperature optimization, peak temperature is reduced effectively, while the temperature variance is significantly lessened. After the combination of power management method, the system throughput acheived improvement up to 25% compared to the simply use task mapping algorithm. Moreover, the optimization on thermal equilibrium of the many-core system is significant. The temperature variance value has reduced by 8 % to 55% compared with the circumstance of TWeNA only. The energy utilization efficiency of the system has also obtained a nearly 15% increase under TDP constraint. |
PDF Full Text Request