LofoSwitch: An online policy for concerted server and disk power control in content distribution networks

doi:10.1016/j.adhoc.2014.11.006

Ad Hoc Networks

Volume 25, Part B, February 2015, Pages 606-621

https://doi.org/10.1016/j.adhoc.2014.11.006 Get rights and content

Abstract

The global content distribution infrastructure is expanding towards the edge of the Internet with cache servers deployed closer to the clients to enable a highly available and bandwidth-efficient delivery of high-quality video services to a broadening range of consumer terminals. The proliferation of cache servers in telco data centers contributes to the unsustainable increase of data-center power consumption. In this paper, we propose to save energy in content distribution networks (CDNs) by means of combined dynamic power management (DPM) of the CDN’s cache servers and the disks in these servers. We present an online load-directed threshold-based power control policy that reduces the CDN’s power consumption while ensuring the best possible viewing experience and a high service availability. In addition, the proposed policy avoids an increase of the bandwidth cost and an excessive wear of the servers and disks. We compare such online policy with a near-optimal offline heuristic greedy policy that determines an upper bound for the energy savings that can be realized by DPM. The evaluation of the power control policies is based on a CDN energy simulator driven by HTTP adaptive-streaming workload traces recorded by an operational telco CDN delivering IPTV services to mobile devices. Even for a minimally provisioned CDN, the online policy reduces the cache-server power consumption by 24%. The energy savings realized by the online policy correspond to 85% of the energy-savings upper bound determined by the offline policy.

Introduction

Between 2011 and 2016, the global IP traffic is forecast to more than triple from 30 to 110 exabytes per month [1]. By 2016, 86% of the global consumer traffic is expected to be a form of video. The exponential growth of online video services is spurred by the massive adoption of smartphones and tablets, which enable end users to watch videos anywhere anytime. According to [1], the IP traffic originating with smartphones and tablets will more than double every year in the period from 2011 till 2016. Growing end user demand for premium video content on mobile devices represents an opportunity for Internet service providers (ISPs) to extend their IPTV offering with online multi-screen video services. ISPs can guarantee a high-quality viewing experience by streaming the video content from cache servers deployed in their own regional networks. Such cache servers constitute a so-called telco CDN. However, the deployment of these servers contributes to the global increase of data-center power consumption, especially because the current generation of CDN infrastructure was not designed for energy efficiency. Koomey [2] estimates that the power consumed worldwide in data centers increased by 50% from 2005 until 2010. In 2010, data centers accounted for already up to 1.5% of the global electricity use. Such exponential growth of data-center power consumption is unsustainable.

Over the last decade, many power-reduction techniques have been proposed for data-center storage systems because such systems account for 25–35% [3] of data-center power consumption. Data-center storage systems draw much power because they consist primarily of hard disk drives, which require power to keep their platters spinning. Also cache servers in content distribution networks contain many disks. Such servers are equipped with disks for caching video-on-demand (VOD) files. Therefore, some of the power-reduction techniques devised for data-center storage systems may be applicable to content distribution networks. Our survey on power-reduction techniques for data-center storage systems [4] reveals dynamic power management (DPM) as the basis for most of these techniques because DPM has the potential of making a system energy proportional, which means that the system’s power consumption is proportional to its workload [5]. DPM turns off system components when they are underutilized to reduce idle power consumption. In this paper, we propose to apply DPM in content distribution networks at the level of the cache servers and the disk drives contained in such servers. We target minimal energy dissipation by turning off underutilized cache servers and over-performing disks. The problem to be solved is when to power which servers and disks on and off. We specifically look for an online power-control policy. Such policy decides on power-state transitions without any knowledge about the future workload. In [6], we present a near-optimal offline power-control policy that we use in this paper to benchmark our proposed online policy. The online power-control policy needs to adhere to the following five constraints.

1.
Unaffected viewing experience: We do not allow DPM to affect the CDN’s global load balancing, i.e., the distribution of the workload across the CDN’s data centers. The global load-balancing policy commonly redirects every client request to the CDN data center that can provide the best possible viewing experience to the client.
2.
High availability: The CDN’s availability is defined as the percentage of requested bytes that can be delivered. This definition matches the definition of service availability proposed by Mathew et al. [7]. The availability needs to exceed the required level.
3.
High bandwidth efficiency: We define the CDN’s bandwidth efficiency as the percentage of uploaded bytes that can be transported from the origin server to the caches at a rate below a predetermined limit. This definition of the bandwidth efficiency is similar to the definition of the availability. The bandwidth efficiency needs to stay above the required level.
4.
No excessive device wear: For every cache server and disk, the number of power-state transitions per day needs to stay below a maximum to avoid a shortened device life span.
5.
Limited time complexity: Because we target an online power-control policy, the policy’s time complexity needs to be limited such that we can ignore the time it takes to decide which servers and disks to turn on or off.

The constraint on bandwidth efficiency ensures that the bandwidth cost does not increase when applying DPM. In practice, the bandwidth-efficiency constraint can easily be tailored to the applicable bandwidth-cost model.

In this paper, we make the following research contributions. As our main contribution, we propose a load-directed threshold-based online power-control policy for two-level dynamic power management in content distribution networks. The two levels addressed by this power-control policy are (1) the level of the cache servers and (2) the level of the disks contained in these servers. At the start of every new time interval, this policy determines which cache servers and disks to power up or down. Our policy is inspired by the Hibernate policy proposed by Mathew et al. [7]. Hibernate is an online policy for one-level DPM in content distribution networks. This one-level power-control policy only affects the cache servers, not the disks. Although Hibernate is a server power-control policy, its inventors were inspired by the threshold-based disk spin-down policy widely used in laptops [8]. We extend Hibernate towards two-level DPM, which also affects the disks contained in the cache servers. We call our policy LofoSwitch because it relies on a last-off first-on (LOFO) policy for deciding which servers and disks to power up and a last-on first-off (LOFO) policy for determining which of the idle servers to deactivate. LofoSwitch allows trading off disk power consumption against the CDN’s bandwidth efficiency in addition to the trade off between server power consumption and the CDN’s availability already enabled by Hibernate. While an active cache server contributes to the CDN’s availability by means of a constant known delivery capacity, an active disk contributes to the CDN’s bandwidth efficiency by means of a time-varying unknown capacity to reduce the upload rate, which is the data rate from the origin server to the cache servers. The decrease in upload rate provided by an active disk is unknown and time-varying because such decrease depends on the disk’s cache efficiency, which varies over time and is unknown in advance. Therefore, the extension from Hibernate to LofoSwitch is not trivial. We evaluate LofoSwitch using the CDN energy simulator we described in [9] and compare this online policy with the near-optimal offline policy we presented in [6].

As a supporting contribution, we propose a policy for distributing the workload over the cache servers in a data center that separates the linear-video from the VOD traffic. In addition, the load is concentrated on as few servers as possible. Such per-stream-type load-concentration policy maximizes the cache efficiency of servers and disks thanks to increased file sharing. This increase in caching efficiency may result in additional energy savings when applying DPM. In addition, cache servers that exclusively handle requests for linear-video files do not need to be equipped with disks because only VOD content benefits from disk caching. Therefore, this load-concentration policy can reduce the hardware cost. Our CDN energy simulator supports this proposed load-concentration policy, which we use to evaluate LofoSwitch. Note that load concentration is a synonym for load consolidation.

The remainder of this paper has the following outline. Section 2 reviews our trace-driven CDN energy simulator and introduces our novel per-stream-type load-concentration policy as part of the CDN’s request-routing system. In Section 3, we briefly revisit our offline near-optimal greedy heuristic power-control policy used to benchmark our novel online power-control policy LofoSwitch. In Section 4, we present our online load-directed threshold-based power-control policy LofoSwitch. This online policy is evaluated in Section 5. Section 6 describes the related work. We conclude our paper in Section 7.

Section snippets

CDN energy simulator

In this section, we briefly review the trace-driven CDN energy simulator that we proposed in [9] and describe a new local load-balancing policy that improves the cache efficiency of the disks by separating linear video from VOD. While in [9] we ignore power-state transition costs, in this paper we take such costs into account. This CDN energy simulator was developed specifically to evaluate power-control policies for two-level DPM, i.e., the combination of cache-server and disk power

Offline power-control policy

In this section, we briefly review the offline near-optimal greedy heuristic power-control policy we presented in [6]. This offline policy is a source of inspiration for the online power-control policy LofoSwitch, which we propose in Section 4. In addition, we use the offline power-control policy in Section 5 to evaluate LofoSwitch.

This offline power-control policy targets a solution to the discrete constrained optimization problem (3). $\begin{matrix} {minimize}_{p_{j}} & E_{j} (p_{j}) \\ subject to & D_{\max, j} (p_{j}) ⩾ D_{req, j} \\ U_{req, j} (p_{j}) ⩽ U_{\max} \end{matrix}$

Online power-control policy

In this section, we present LofoSwitch, an online load-directed threshold-based power-control policy for two-level dynamic power management in content distribution networks. This power-control policy extends the online power-control policy for one-level DPM called Hibernate [7] towards two-level DPM. Similar to Hibernate, LofoSwitch uses a threshold to determine when to power devices down. Such threshold-based policy is inspired by the threshold-based disk spin-down policy commonly used as

Evaluation

In this section, we evaluate LofoSwitch, the proposed threshold-based load-directed online power-control policy for two-level DPM in CDNs presented in Section 4. We benchmark LofoSwitch against the near-optimal heuristic greedy offline power-control policy for two-level DPM in CDNs reviewed in Section 3. In addition, we compare LofoSwitch with the original Hibernate for one-level DPM in CDNs proposed in [7].

The evaluation is based on the trace-driven CDN energy simulator described in Section 2.

Related work

This paper is the first to combine server and disk power management in content distribution networks. Therefore, we build on the work done in two research domains that have developed so far rather independently: energy-aware data storage systems on the one hand and energy-efficient content distribution on the other hand.

Based on an analysis of over a hundred high-quality papers published over the last decade, we wrote a comprehensive survey on power-reduction techniques for data-center storage

Conclusion

The increasing demand for multiscreen IPTV services spurs the deployment of telco content distribution networks. The roll out of cache servers in data centers close to the IPTV subscribers fuels the data-center power consumption. Such cache servers are typically filled with disks to support video-on-demand. Therefore, we propose saving energy in telco CDNs by applying dynamic power management to the cache servers and their disks. We present LofoSwitch, an online load-directed threshold-based

Acknowledgments

The authors would like to thank their colleagues of Velocix, an Alcatel-Lucent company, and Koen Laevens for their support in getting access to CDN workload traces. We also thank Bart De Vleeschauwer and the anonymous reviewers for making numerous valuable suggestions for improvement. In addition, this work is supported by the Flanders Agency for Innovation by Science and Technology (IWT), Grant IWT 100690.

Tom Bostoen obtained his B.Sc. (1995) and M.Sc. (1998) degrees in Physics Engineering from Ghent University in Belgium. Since then, he has been with Alcatel-Lucent in Antwerp, Belgium, where currently, he works as a researcher in Bell Labs pursuing a Ph.D. in Computer Science in collaboration with the KU Leuven in Belgium. His research interests include energy-efficient networking, content distribution networks, customer experience management, and access networks. Prior to his current

References (44)

Cisco Visual Networking Index: Forecast and Methodology, 2012–2017, 2013...
J. Koomey, Growth in Data Center Electricity Use 2005 to 2010, Tech. Rep., Analytics Press, Oakland, CA, USA...
J. Kim et al.
Energy proportionality for disk storage using replication
T. Bostoen et al.
Power-reduction techniques for data-center storage systems
ACM Comput. Surv.
(2013)
L.A. Barroso et al.
The case for energy-proportional computing
Computer
(2007)
T. Bostoen et al.
Minimizing energy dissipation in content distribution networks using dynamic power management
V. Mathew et al.
Energy-aware load balancing in content delivery networks
T. Bostoen et al.
Analysis of disk power management for data-center storage systems
T. Bostoen et al.
A simulator to assess energy-saving techniques in content distribution networks
K. Stamos et al.
Cdnsim: a simulation tool for content distribution networks
ACM Trans. Model. Comput. Simul.
(2010)

M. Cha et al.

Watching television over an ip network

P. Gill et al.

Youtube traffic characterization: a view from the edge

H. Kellerer et al.

Knapsack Problems

(2004)

A. Qureshi et al.

Cutting the electric bill for internet-scale systems

E. Pinheiro et al.

Exploiting redundancy to conserve energy in storage systems

SIGMETRICS Perform. Eval. Rev.

(2006)

D. Li et al.

Eeraid: energy efficient redundant and inexpensive disk array

X. Yao et al.

Rimac: a novel redundancy-based hierarchical cache architecture for energy efficient, high performance storage systems

SIGOPS Oper. Syst. Rev.

(2006)

L. Lu et al.

Diskgroup: energy efficient disk layout for raid1 systems

J. Wang et al.

eraid: conserving energy in conventional disk-based raid system

IEEE Trans. Comput.

(2008)

K.M. Greenan et al.

A spin-up saved is energy earned: achieving power-efficient, erasure-coded storage

D. Harnik et al.

Low power mode in cloud storage systems

J. Leverich et al.

On the energy (in)efficiency of hadoop clusters

SIGOPS Oper. Syst. Rev.

(2010)

Cited by (6)

A cloud server energy consumption measurement system for heterogeneous cloud environments
2018, Information Sciences
With rapid development of cloud computing technologies and applications, the number and scale of cloud data centers have grown exponentially in recent years. One of the major problems with current cloud data centers is their huge energy consumption, which makes energy consumption management one of the hottest research topics in the field of cloud computing. This paper aims at implementing an effective Distributed Energy Meter (DEM) system for heterogeneous cloud environments based on a multi-component power consumption model for cloud servers. Specifically, we propose a modeling method for the energy consumption of key components (CPU, memory and disk) of computer servers and reveal the mathematical relationship between the resource usage of the key components and the system energy consumption. The proposed DEM system cannot only estimate the energy consumption of heterogeneous cluster environments (Linux and Windows NT), but also support various CPU power consumption models. In addition, a unique disk power consumption model that uses different thresholds to distinguish various disk I/O states (sequential/random, read/write) to achieve an accurate estimation of disk power consumption. Experimental studies conducted on a heterogeneous cluster with workloads generated by PCMark and Sysbench demonstrate that the proposed DEM system outperforms the state-of-art models in estimating the energy consumption of heterogeneous cloud environments.
Experimental and quantitative analysis of server power model for cloud data centers
2018, Future Generation Computer Systems
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.
Dynamic analysis of access control based on partially observable markov decision process for media distribution network
2019, 2019 4th IEEE International Conference on Advanced Robotics and Mechatronics, ICARM 2019
A Distributed Cache Update Deployment Strategy in CDN
2018, Journal of Physics: Conference Series
Energy-optimal collaborative file distribution in wired networks
2017, Peer-to-Peer Networking and Applications
Energy analysis for dynamic cache storage in video on demand services
2016, 2016 21st OptoElectronics and Communications Conference, OECC 2016 - Held Jointly with 2016 International Conference on Photonics in Switching, PS 2016

Sape Mullender is chief cook and bottle washer for the wireless systems research in Alcatel-Lucent’s Bell Laboratories and a part-time extraordinary professor of computer science at the University of Twente in the Netherlands. His research interests include operating systems, multimedia systems, and wireless systems. He received his Ph.D. from the Vrije Universiteit Amsterdam and was a faculty member there until 1983. From 1984 to 1990, he headed the distributed systems and computer networks research group at the Centre of Mathematics and Computer Science (CWI) in Amsterdam. From 1991 to 1998, he was a full professor at the University of Twente. He started work at Bell Labs in 1998.

Yolande Berbers obtained her M.Sc. (1982) and Ph.D. (1987) degrees in Computer Science from the KU Leuven (Belgium). Since 1990, she has been a professor at the department of computer science. Her research interests include ubiquitous computing, context-aware computing, service architectures, middleware, cloud computing, software engineering for embedded software, distributed systems, and mobile agents. She teaches courses on software for real-time and embedded systems and on computer architecture. Currently, she is also vice-dean of KU Leuven’s Faculty of Engineering and president of the Leuven Center on Information and Communication Technology (LICT) of the KU Leuven.

View full text

LofoSwitch: An online policy for concerted server and disk power control in content distribution networks

Abstract

Introduction

Section snippets

CDN energy simulator

Offline power-control policy

Online power-control policy

Evaluation

Related work

Conclusion

Acknowledgments

Energy proportionality for disk storage using replication

Power-reduction techniques for data-center storage systems

ACM Comput. Surv.

The case for energy-proportional computing

Computer

Minimizing energy dissipation in content distribution networks using dynamic power management

Energy-aware load balancing in content delivery networks

Analysis of disk power management for data-center storage systems

A simulator to assess energy-saving techniques in content distribution networks

Cdnsim: a simulation tool for content distribution networks

ACM Trans. Model. Comput. Simul.

Watching television over an ip network

Youtube traffic characterization: a view from the edge

Knapsack Problems

Cutting the electric bill for internet-scale systems

Exploiting redundancy to conserve energy in storage systems

SIGMETRICS Perform. Eval. Rev.

Eeraid: energy efficient redundant and inexpensive disk array

Rimac: a novel redundancy-based hierarchical cache architecture for energy efficient, high performance storage systems

SIGOPS Oper. Syst. Rev.

Diskgroup: energy efficient disk layout for raid1 systems

eraid: conserving energy in conventional disk-based raid system

IEEE Trans. Comput.

A spin-up saved is energy earned: achieving power-efficient, erasure-coded storage

Low power mode in cloud storage systems

On the energy (in)efficiency of hadoop clusters

SIGOPS Oper. Syst. Rev.