Power Model And Monitoring System For Heterogeneous Cloud Servers

With the rapid development of cloud computing technology currently,the scale of cloud data centers is also increasing,which brings about huge energy consumption.Based on this background,the energy consumption management and optimization of cloud data center has become an important research direction of data center management,and the power model and power monitoring for heterogeneous cloud servers are the basis of energy consumption research.In order to solve the problem that the existing CPU power models are not accurate enough and do not fully consider the heterogeneity of CPU,this paper proposes new CPU power models based on CPU architectures x86 and ARM.On the x86 architecture,this paper proposes a new power-exponential function CPU power model.Compared with the existing linear model,polynomial model and power function model,this model provides a higher accuracy in the CPU power measurement of the latest servers.This paper also improves the model(IPEFM)by considering the multi-threading and DVFS characteristics of modern CPU to improve the model accuracy.On the ARM architecture,this paper proposes a new CPU power model based on hybrid modeling(HBM).This model takes into account the modeling features of CPU utilization and CPU performance events,and requires only a small cost of model training to achieve high accuracy.In this paper,the relationship between I/O mode and utilization and disk power is further studied,and a disk power model combining I/O mode and utilization is proposed(MUBM).The I/O mode of disk is distinguished by the corresponding thresholds,and I/O utilization is taken as one of the modeling parameters,so as to improve the measurement accuracy of disk power.On the other side,this paper designs and implements a distributed power monitoring system for heterogeneous cloud servers(DPMS).DPMS uses an improved hardware-aware CPU power measurement(IHCPM)on the CPU power measurement of x86 architecture.IHCPM automatically selects the appropriate CPU power model based on the CPU launch year and does not require model training cost.DPMS can measure the power of server components and the whole machine,virtual machine and container based on the proposed power models and power measurement.DPMS can be deployed on both Windows and Linux platforms,and supports CPU power measurement for x86 and ARM architectures.DPMS also adopts an efficient distributed communication and storage architecture that can adapt to complex and heterogeneous cloud environments.At last,this paper uses workload tools such as PC Mark and Sysbench to verify the proposed CPU power models,disk power model and IHCPM,and the mean relative errors(MREs)of the models and measurement are within 4%,having good accuracy.At the same time,it also uses customized workload scripts to conduct distributed cluster,virtual machine and container power experiments on DPMS to verify the effectiveness of the system.Experimental results show that DPMS can achieve accurate power measurement of heterogeneous cloud servers(MREs are between 1% and 7%),and it is suitable for distributed heterogeneous cloud environment.