Resource Management Algorithm For Multi-core System In Dark Silicon Era

With the prosperous development of VLSI design,the transistor's feature size has been sharply decreasing,which results in higher level of chip integration with theoret-ically constant on-chip power density.But when the technology nodes reach 22 nm,leakage power begins dominating with exponential growth rate which leads to unavoid-able rising of power density and causes the chip thermal problem.Consequently,the concept of dark silicon design is proposed to avoid the hotspots,where all the fractions of chip resources cannot be simultaneously powered on.It reveals the coming of the dark silicon era.The multi-core system emerged due to the performance limitation of single-core system,and the trend of dark silicon require the cores to be powered on and configured properly to alleviate heat problem.Dark silicon derives a series of challenges from vari-ous perspectives in circuit design,including system architecture and dark core accelera-tor,low power techniques,resources management methods,etc.Among these research areas,resources management methods can reasonably allocate thermal design power,frequency level and the number of processors to maximize the system performance ac-cording to features of the input applications within thermal resources constrains.This thesis proposes a resources management method for the multi-core system with symmetric shared memory architecture in dark silicon design.Firstly,given a set of applications with diverse degrees of thread level parallelism,we extract the throughput and power consumption of each application under all the possible processor number and frequency configuration.Constrained by the system resources and power budget,the resources configuration problem is solved by dynamic programming.Secondly,a simulated annealing based method is used to map the applications and determine the distribution of dark and active cores to minimize thermal and communication costs.Finally,according to the feedback that whether hotspots exist in the system,a iterative power adaption method is used to obtain maximum power budget and avoid temperature violation.We build the experiment framework to analyse the performance and simulate the power and temperature.Supported by the most prevalent simulation tools,we redact the script files as the bridge to connect all tools and the format converting code that makes sure the proposed algorithm can be embedded into the framework and fluently executed.This framework is applicable for basic homogenous multi-core system which can obtain performance analysis for given application sets,power consumption of each system component and temperature distribution result.Experimental results show that,compared with chess mapping approach,the maximum temperature decreases about 3%at best.Additionally,we obtain 12%gain in performance when compared with power down hotspot adaption.