Research On Power Budget Of Homogeneous And Heterogeneous Dark Silicon System

With the continuous improvement of the technology level,the power density of devices has been increasing,resulting in the high temperature of current highperformance multi-core processors.Excessive temperature will damage the chip,so there are many thermal related problems,such as system reliability problems and performance degradation problems.At the same time,due to the increase in device power density,in order to ensure that the chip temperature is below a safe temperature,the core of a multi-core chip cannot run at high frequencies at the same time,so some cores have to be shut down.Traditional dynamic thermal management has not considered dark silicon systems.For current multicore processors,the ratio of unopened cores to the total number of cores exceeds 50%.What's more serious is that as the process improves,the static power consumption will also increase,making the dark silicon problem more serious.When using the 8nm process,the dark silicon area can reach up to 80% of the total area,which greatly reduces the chip performance.So dark silicon systems are a major limiting factor in the performance of modern multicore chip systems.To solve this problem,power budget-constrained dynamic optimization is performed in various ways,including dynamic voltage and frequency scaling(DVFS)and task scheduling techniques.However,the power budget given by some methods is usually too pessimistic,which greatly limits the ability of dynamic performance optimization methods.In this paper,we propose a thermal management scheduling algorithm called GDP(Greedy based Dynamic Power)algorithm and PPB(parallel power budgeting)algorithm for homogeneous systems and heterogeneous systems to maximize system performance.Both of these algorithms can dynamically give the chip's open core position and core power to maximize chip performance.The GDP algorithm transforms the problem of maximizing system performance into an optimization problem of maximizing the average temperature of the system.Combined with the basic idea of the greedy algorithm and the orthogonal matching pursuit algorithm,the open-core position of the current chip is found in turn.The PPB algorithm transforms the problem of maximizing system performance into a power optimization problem under constraints.At the same time,the PPB algorithm is also based on the idea of greed,and gradually determines the position of the core opening.Based on this idea,both algorithms can find a suboptimal solution for the distribution of open core positions,which improves system performance.Both the GDP algorithm and the PPB algorithm can consider transient thermal effects and current temperature effects,which is not possible with other methods.These two algorithms provide a higher power budget and guarantee the thermal safety of the chip under the premise of low time cost,making it superior to other tasks.In short,the GDP algorithm and PPB algorithm reduce the time complexity,realize dynamic management,and can automatically determine the core opening position,calculate the power budget value,and automatically improve the chip performance based on the chip's current temperature.