Elsevier

Future Generation Computer Systems

Volume 86, September 2018, Pages 940-950
Future Generation Computer Systems

Experimental and quantitative analysis of server power model for cloud data centers

https://doi.org/10.1016/j.future.2016.11.034Get rights and content

Highlights

  • Modeling server power at component-level.

  • A systematic review on widely-used component power models.

  • Quantitative analysis based on SPEC data and experimental evaluations.

  • Research hotspots and prospects related to power model.

Abstract

Scientific computing applications like online social network analysis demand enormous computing capability from cloud service, but now the high energy consumption by cloud data centers has brought more concerns on power monitoring and management to cloud service providers (CSPs). Compared with hardware-based traditional techniques, server power monitoring based on power model is of higher scalability as well as lower deployment cost and thus, is more feasible for cloud data center power management. However, previous studies lack a systematic review and quantitative analysis on server power model. In this paper, we review and compare several popular power models of cloud server components including CPU, vCPU, memory and hard disk. We propose an I/O-mode aware disk power model based on our observation of disk power behavior. Experimentally, we first analyze the accuracy of different CPU power models by looking into a SPECpower_ssj2008 dataset. We also carried out experiments on a physical server to evaluate memory power models and disk power models. The experimental results indicate the advantage of polynomial CPU model, LLCM-based memory model and the proposed disk model. The ideology of component-level power modeling presented in this paper helps realize fine-grained power control. Moreover, the evaluation and comparison results provide CSPs with useful guidance on optimizing energy management of cloud data centers.

Introduction

As the demands for scientific computing become more and more common, cloud computing is quickly gaining its popularity. According to statistics, globally there were already over 500,000 data centers, 2011  [1]. Cloud infrastructures provide enormous and elastic computing capability for big-data applications like online social network (OSN) analysis. However, a huge number of users are active everyday making some OSN applications incredibly energy-consuming  [2]. A report shows that data centers account for more than 1.5% of the global electricity consumption  [3]. For service providers (CSPs), it is critical to reduce energy cost. Effective power monitoring is the foundation of constructing a “green” data center. In cloud environment, the monitoring system should be highly scalable and able to provide accurate and fine-grained power data. An Emerson’s report across North America shows that 51% of the surveyed respondents cited adequate monitoring/data center management capabilities among their three biggest concerns  [4]. Thus it is an urgent need to establish such monitoring mechanism.

Traditional power measuring techniques reply on physical meters or hardware sensors to collect server power. These methods are feasible in small-scale data centers. However, usually they are not applicable in large-scale and heterogeneous cloud data centers because it is too costly to install a meter or a data acquisition (DAq) system (e.g., IBM Active Energy Manager APIs  [5]) on every server. Besides, it is tough to resolve the problems of bad compatibility and low scalability. Nowadays, power model is widely used as a method to estimate server power and it also provides a means for monitoring virtual machine’s power as well. For example, Joulemeter  [6] was developed on the basis of a VM power model proposed in  [7]. Some studies on power management and resource scheduling (e.g.,  [8], [9], [10]) validated their strategies or algorithms by means of simulation. As a matter of fact, most simulation tools like CloudSim  [11] exploit built-in power models to simulate server’s power behavior.

Most of the relevant studies built models at the level of physical server but actually power estimate can be done in a more fine-grained manner. Researchers used to figure out that CPU and memory are the major power-consuming components  [12], [13]. However, more and more data-intensive jobs are submitted to the cloud as big-data processing gradually becomes a common need. So the power consumption of storage devices is not negligible in many cases. In this paper, we review and evaluate widely-used power models of CPU, memory and hard disk. A reconfiguration-adaptive vCPU model is also presented. Firstly, we built a dataset by collecting power data from SPECpower_ssj2008. The dataset contains power data measured on hundreds of different models of servers, which were mainly reported by server manufacturers from the 4th quarter of 2007 to the 1st quarter of 2016 (totally 34 quarters). We used these data to fit different forms of CPU power models and calculated their error distributions. The result is presented in the form of cumulative distribution function. Secondly, we discuss two types of memory power models. One is based on performance counters and the other one is based on memory usage. By analyzing experimental results, we show that the correlation between memory power and LLC misses is much stronger than the one between memory power and memory usage. We also evaluated the accuracy of LLCM-based memory power model. Thirdly, we studied the power behavior of hard disk in different I/O modes through running I/O-intensive benchmarks. We propose an I/O-mode aware disk power model based on our observation and proved its accuracy experimentally. The study on server power model is of great significance to improve power monitoring and management in cloud data centers. We also discuss the application of server power model in some major research trends at the end of this paper.

Section  1 briefly introduces the background and major contributions of our study. In Section  2, we review widely-used component power models, present a vCPU power model and propose a novel I/O-mode aware disk power model. We demonstrate the evaluation results of CPU, memory and disk power models in Section  3. Section  4 is about the discussion on power model’s application in future research. Finally, we make conclusions in Section  5.

Section snippets

Server power model

Basmadjian et al.  [14] categorized the ICT infrastructures of a data center into servers, network devices, cooling system and power supply units. According to statistics data, servers and storage nodes consume 45% of the data center’s total power  [15]. So in this paper, we mainly focus on modeling server power by breaking it down into components. As mentioned in  [7], CPU, memory and disk are the major components consuming most of the system’s power. We discuss widely-used component power

Methodologies and experimental setup

We mainly study the dynamic power of CPU, memory and disk under specified workload. That means their idle power is counted together: Pidle=Pcpu_idle+Pmem_idle+Pdisk_idle+Pothers wherePcpu_idle,Pmem_idle andPdisk_idle denote the idle power of CPU, memory and disk, respectively.Pothers is the power consumption of other components. In practice, it is very difficult to measure the idle power of each component separately. Thus, we consider the idle power (Pidle) at system-level. Experimentally, we

Power modeling for VMs and containers

Virtualization is a prevalent technique applied to physical servers in cloud data centers. Hardware virtualization plays an important role in maximizing resource utilization but at the same time poses many challenges to cloud data centers’ power monitoring. Kansal et al.  [7] proposed a multi-components virtual machine power model and developed its implementation named Joulemeter  [6]. However, Joulemeter relies on the APIs provided by Windows Management Instrumentation (WMI) and Hyper-V to

Conclusions

The rapid growth of cloud computing’s popularity are bringing more and more concerns on power monitoring and management to service providers. Server power model is widely used in cloud data center power monitoring and resource scheduling evaluation. We introduce a methodology to estimate server power and review popular power models of CPU, memory and hard disk. Apart from that, a vCPU power model adaptive to reconfiguration is also presented. We collected a large dataset from SPECpower_ssj2008

Acknowledgments

Thanks to the helpful comments and suggestions from the anonymous reviewers. This work is partially supported by the National Natural Science Foundation of China (Grant No. 61402183), Guangdong Natural Science Foundation (Grant Nos. S2012030006242 and S2013040012449), Guangdong Provincial Scientific and Technological Projects (Grant Nos. 2016A010101007, 2016B090918021, 2014B010117001, 2014A010103022 and 2014A010103008), Guangzhou Science and Technology Projects (Grant Nos. 201607010048 and

Weiwei Lin received his B.S. and M.S. degrees from Nanchang University in 2001 and 2004, respectively, and received his Ph.D. degree in Computer Application from South China University of Technology in 2007. He is currently an associate professor in the School of Computer Science and Engineering, South China University of Technology. His research interests include distributed system, cloud computing and big data.

References (67)

  • A. Mishra et al.

    Energy efficient voltage scheduling for multi-core processors with software controlled dynamic voltage scaling

    Appl. Math. Model.

    (2014)
  • X. Li et al.

    Energy efficient virtual machine placement algorithm with balanced and improved resource utilization in a data center

    Math. Comput. Model.

    (2013)
  • State of the Data Center 2011. Available:...
  • F. Jalali, C. Gray, A. Vishwanath, R. Ayre, T. Alpcan, K. Hinton, R.S. Tucker, Energy consumption of photo sharing in...
  • J. Koomey

    Growth in Data Center Electricity Use 2005 to 2010, Technical Report

    (2011)
  • Emerson Network Power Survey of North American Data Center Users’ Group, 2014. Available:...
  • A. Verma, P. Ahuja, A. Neogi, pMapper: Power and migration cost aware application placement in virtualized systems, in:...
  • Microsoft Joulemeter. Available:...
  • A. Kansal, F. Zhao, J. Liu, N. Kothari, A.A. Bhattachrya, Virtual machine power metering and provisioning, in: ACM...
  • W.W. Lin et al.

    Novel algorithms and equivalence optimization for resource allocation in cloud computing

    Int. J. Web Grid Serv.

    (2015)
  • CloudSim. Available:...
  • J.C. Mccullough, Y. Agarwal, J. Chandrashekar, S. Kuppuswamy, A.C. Snoeren, R.K. Gupta, Evaluating the effectiveness of...
  • R. Basmadjian, N. Ali, F. Niedermeier, H.D. Meer, G. Giuliani, A methodology to predict the power consumption of...
  • C.H. Hsu, S.W. Poole, Power signature analysis of the SPECpower_ssj2008 benchmark, in: IEEE International Symposium on...
  • A.E.H. Bohra, V. Chaudhary, VMeter: Power modelling for virtualized clouds, in: Proceedings of IEEE International...
  • T.N. Mudge

    Power: a first-class architectural design constraint

    Computer

    (2001)
  • K.J. Ye et al.

    Power management of virtualized cloud computing platform

    Chinese J. Comput.

    (2012)
  • Y.C. Lee et al.

    Energy efficient utilization of resources in cloud computing systems

    J. Supercomput.

    (2010)
  • X. Fan et al.

    Power provisioning for a warehouse-sized computer

    ACM SIGARCH Comput. Archit. News

    (2007)
  • R. Buyya, A. Beloglazov, J. Abawajy, Energy-efficient management of data center resources for cloud computing: A...
  • J. Song et al.

    Study on energy-consumption regularities of cloud computing systems by a novel evaluation model

    Computing

    (2013)
  • S. Rivoire, P. Ranganathan, C. Kozyrakis, A comparison of high-level full-system power models, in: The Workshop on...
  • W.T. Wu et al.

    An intelligent power consumption model for virtual machines under CPU-intensive workload in cloud environment

    Soft Comput.

    (2016)
  • Cited by (29)

    • EAIS: Energy-aware adaptive scheduling for CNN inference on high-performance GPUs

      2022, Future Generation Computer Systems
      Citation Excerpt :

      The energy consumption of GPUs is the most important part of inference service energy-efficiency optimization. Besides, previous studies [48–51] have explored solutions to the energy consumption of services and network systems. The above techniques are complementary to EAIS to achieve better results.

    • Workload forecasting and energy state estimation in cloud data centres: ML-centric approach

      2022, Future Generation Computer Systems
      Citation Excerpt :

      The data-driven methods based on machine learning are being researched to save energy and optimize resource usage. The parameters like CPU, memory, disk and idle power all contribute to a host’s total energy [8]. To accurately estimate the energy consumption of the host, all the necessary contributing elements should be considered.

    • Rural smart tourism under the background of internet plus

      2021, Ecological Informatics
      Citation Excerpt :

      The system can then transmit an alert, which will cause the desired reaction. This developing technology has the potential to provide high-value-added use applications, such as identifying or forecasting crimes (Lin et al., 2018). The intelligent monitoring management system includes management module, device access module, alarm management module, Web service module, streaming media forwarding and management module, storage module, mobile access module, client access module, intelligent analysis service module, etc. (Abad and Álvarez, 2020).

    • A cloud resource management framework for multiple online scientific workflows using cooperative reinforcement learning agents

      2020, Computer Networks
      Citation Excerpt :

      With the growing number of users and applications and their need for systems with powerful resources, enterprise distributed and parallel networks were created [1]. Cloud is one of the most popular distributed environments, in which, powerful data resources called datacenters, which include powerful processors, large memory space, and powerful connectivity networks, are available for cloud users [2,3]. The Amazon web service (AWS),1 Microsoft Azure2 and the Google cloud platform (GCP)3 are some of the most important cloud service providers, resources of which are shared by users' applications.

    View all citing articles on Scopus

    Weiwei Lin received his B.S. and M.S. degrees from Nanchang University in 2001 and 2004, respectively, and received his Ph.D. degree in Computer Application from South China University of Technology in 2007. He is currently an associate professor in the School of Computer Science and Engineering, South China University of Technology. His research interests include distributed system, cloud computing and big data.

    Wentai Wu is a master student in Computer Science at South China University of Technology. He received his bachelor degree from South China University of Technology in 2015. His research interests include cloud computing and data center power management.

    Haoyu Wang is currently pursuing his master degree in Computer Science in the School of Computer Science and Engineering, South China University of Technology. His research interests are mainly related to distributed computing and cloud resource management.

    James Z. Wang received his B.S. and M.S degrees in Computer Science from University of Science and Technology of China. He obtained his Ph.D. degree in Computer Science from University of Central Florida. He is currently a professor in the School of Computing at Clemson University, South Carolina. His research includes storage network, database system, distributed system, cloud computing and multimedia technologies. Dr. Wang is a senior member of IEEE and ACM.

    Ching-Hsien Hsu is a professor in the Department of Computer Science and Information Engineering at Chung Hua University, Taiwan. His research includes high performance computing, cloud computing, big data intelligence, parallel and distributed systems, ubiquitous/pervasive computing and intelligence. Dr. Hsu is an IEEE senior member.

    View full text