An energy-saving scheduling scheme for streaming media storage systems①

Shang Qiuli (尚秋里)

(National Network New Media Engineering Research Center, Institute of Acoustics,Chinese Academy of Sciences, Beijing 100190, P.R.China)(University of Chinese Academy of Sciences, Beijing 100190, P.R.China)

?

An energy-saving scheduling scheme for streaming media storage systems①

Shang Qiuli (尚秋里)②

(National Network New Media Engineering Research Center, Institute of Acoustics,Chinese Academy of Sciences, Beijing 100190, P.R.China)(University of Chinese Academy of Sciences, Beijing 100190, P.R.China)

The rapid growth of streaming media applications on the Internet is proposing higher requirements on energy consumption and I/O performance of the storage systems. However, the optimized I/O requests from different initiators will be mixed disorderly when they are reaching the storage system concurrently, which leads to increasing energy consumption. This paper proposes an energy-saving scheduling scheme based on I/O Stream (ES-IOS). The ES-IOS scheme can take the advantage of the I/O characteristics of streaming media and reorganize the mixed and disordered I/O requests into “streams”. Technically, The ES-IOS scheme includes two main points, a priority-based weighted stream scheduling algorithm (PWSS) and a regression-fitting-based popularity prediction algorithm (RFPP). The PWSS algorithm can schedule the I/O streams in weighted queue based on priority to limit energy consumption. The priority of each stream is determined by its popularity. According to the I/O access records over a period, the RFPP algorithm can predict the popularity of each stream via regression fitting. Based on the popularities, the PWSS algorithm assigns more continuous service time to the hot streams and reversely less service time to the cold ones. Trace-driven experiments show that the ES-IOS scheme can reduce the energy consumption by 38% and enhance the I/O throughput by 27% approximately.

streaming media, energy-saving, I/O characteristic, storage system

0 Introduction

Nowadays, due to the development of network technology and the prevalence of consumer electronics such as smart phones, smart TVs, and tablet computers, streaming media has gradually become the mainstream application on the smart devices. Currently, video streaming constitutes approximately 50% of the current cellular network traffic, which is expected to reach 66% by 2017[1]. The rapid growth of streaming media on the Internet has led to higher requirements on storage systems. The streaming media storage systems are facing increasing challenges in QoS (Quality of Service), I/O performance and energy consumption.

Some studies show that energy consumption of storage subsystem accounts for approximately 50% of the total energy consumption in the computer system[2]. Contemporarily, as the energy price keeps growing, the energy consumption problem of streaming media services has aroused wide concerns in industry and the academia communities. The reason includes the following aspects. First, large-scale enterprises usually spend hundreds of millions of dollars on electric energy consumption of computer systems annually. Second, energy consumption causes higher temperature, which greatly lowers the system reliability. For example, if the operating temperature is 5 degrees Celsius higher than normal, the failure probability of a disk will be increased by 10%～15%[3]. Third, computer systems with higher energy consumption will trigger environmental pollution by generating harmful gas, noise, and electromagnetic radiation.

The storage system for streaming media is always accessed concurrently by multiple initiators. Generally, the I/O requests from each initiator are sorted and merged by the I/O schedulers. However, the optimized I/O requests from different initiators will be mixed disorderly when they reach the storage system concurrently. For the traditional and general storage systems, the mixed I/O requests are simply handled by the FCFS(First Come, First Serve) scheme. Therefore, the mixed I/O requests in these storage systems usually lead to more disk seeks, more read or write penalties, more fragments and checksums to be handled, which will further result in rapid growth of energy consumption and negative effect on I/O throughput performance.

This study proposes an energy-saving scheduling scheme based on I/O stream (ES-IOS). ES-IOS can take the advantage of the I/O characteristics of streaming media and reorganize the mixed and disordered I/O requests into “streams”. Then the streams can be treated as the basic units in scheduling. Technically, ES-IOS includes two main points, a priority-based weighted stream scheduling algorithm (PWSS) and a regression-fitting-based popularity prediction algorithm (RFPP). PWSS can schedule the I/O streams in weighted queue based on priority to limit energy consumption. The priority of each stream is rated by its popularity. According to the I/O access records over a period, RFPP can predict the popularity of each stream via regression fitting. Based on the popularities, PWSS assigns more continuous service time to the hot streams and reversely less service time to the cold ones.

The ES-IOS scheme can effectively reduce the energy consumption and improve the throughput performance of streaming media storage systems via limiting the disk seeks, but may cause some extra access delay.

The rest of this paper is organized as follows. Section 1 briefly outlines the related work on energy saving. Section 2 describes an energy consumption model for streaming media storage systems. Section 3 presents the ES-IOS scheme including the PWSS algorithm and the RFPP algorithm. The experimental evaluations of ES-IOS are described in Section 4. Finally, Section 5 concludes the paper.

1 Related work

1.1 Ideas

Tian, et al.[4]outlined some basic ideas for energy saving in disk storage systems. It can be divided into two aspects as follows.

(1) Decrease the internal energy consumption of the disk. Gurumurthi et al.[5]proposed the concept called DRPM (dynamic revolutions per minute) to modulate disk speed dynamically according to the I/O loads of storage system. To decrease the energy consumption of magnetic head movement, Huang, et al.[6]created copies of the data on some disks in order to redirect the I/O loads to access contiguous addresses as much as possible, which reduced the energy consumption and promoted the throughput performance of the system effectively. In order to extend the spare time of disk, Pathanasiou, et al.[7]cached scattered write operations and processed them in batches.

(2) Limit the number of disks that consuming energy in storage system. This aspect mainly includes PARAID[8], MAID[9]and PDC[10]. Depending on I/O characteristics, they released the I/O loads on some disks via I/O redirection or data redistribution, and further turned off these disks to save energy.

1.2 Methods

Currently, the energy-saving methods are mainly based on two points of view, exploiting the I/O characteristics and adjusting the data distribution on disks.

(1) Exploit the I/O characteristics. Li, et al.[11]proposed eRAID model (energy-efficient RAID) (RAID, redundant array of independent disks) that redirected the I/O requests based on the redundancy of the disks and turned off the rotation of some spare disks to limit energy consumption. Weddle, et al.[8]implemented a self-adaptive organization of disk arrays, called PARAID (power-aware RAID), which can save energy greatly without compromising performance. Zhu, et al.[12]divided the system cache into several parts by different I/O characteristics and assigned these parts to different disk groups dynamically according to the I/O loads and energy consumption. For large-scale archival systems, Colarelli, et al.[9]cached the hot data on extra disks to avoid the back-end disks switching to active mode frequently.

(2) Adjust the data distribution. Based on the dynamic speed disk model, Zhu, et al.[13]migrated some data to disks at appropriate speed to save energy of the storage systems in data center. Pinheiro, et al.[10]tried to move the hot data to some disks periodically in order to keep most of the I/O requests being processed by the hot disks, which can improve the energy efficiency of the system effectively at the cost of sacrificing some system performance. Write offloading technology[14]redirected the write requests to the redundant disk groups to extend the standby time of some disks in large-scale storage systems. In the view of data redundancy in distributed storage systems, diverted access technology[15]redirected the I/O requests to some active storage nodes to prolong the low energy state of other nodes.

In addition, with the development of flash memory, SSD (solid state drive) has been widely used in storage systems. SSD has high random access performance with much lower energy consumption than traditional disks. On one hand, since it makes up the performance gap between DRAM (dynamic random access memory) and disk, SSD can act as the second level cache for I/O requests to reduce energy consumption of active disks. On this aspect, the studies mainly include the EXCES system[16], the BEST technology for mobile devices[17]and the DRAM cache extension through flash memory[18]. On the other hand, the hybrid storage system can be formed by SSD and disk to provide unified access address. The hybrid storage system can redirect the hot I/O requests to SSD to improve the random access performance as well as the energy efficiency. This aspect mainly includes ComboDrive system[19], Hystor system[20]and the data-prefetching scheme for media players proposed by Ryu et al[21].

1.3 Discussion

At present, there have been considerable studies in energy saving problems for general storage systems, but researches focusing on specific scenarios are still far from enough. Some specific applications have obvious I/O features that can be extracted and utilized to guide the system-level optimization of energy saving. Focusing on the application scenario of streaming media, this paper studies the energy-saving schemes for the storage systems via considering the I/O characteristics of streaming media. Besides, because of the limitation of cost and service life of SSD, hardly can SSD drastically replace disks in storage systems at present, especially in streaming media systems. Even in the hybrid storage systems, disk is still the main cause of energy consumption, therefore this paper only takes account of the energy-saving optimization of disk storage systems.

2 Energy consumption model

2.1 Energy model of disk

A typical disk is mainly composed of a number of circular magnetic discs, magnetic heads and motors. The rotation of discs and the seeking of heads are both driven by the motors in the disk. When processing the I/O request, the head will be driven to the exact disc and track, then the data on this track can be moved to the position under the head. Because the motor consumes electric power, the energy consumption of a disk Ediskrests with two parts: the rotation energy Erotateand the seeking energy Eseek.

Edisk=Erotate+Eseek

(1)

As is shown in Fig.1, the energy consumption of disk is generally divided into four modes: active, idle, standby, and sleep. When processing I/O requests, the disk is active with the most energy consumption because the disc rotates at high speed and the head seeks to read and write data. If there is no I/O request temporarily, the disk will be in idle mode with lower energy consumption than the active mode because the disc keeps rotating and the head stops seeking. The disk consumes much lower energy in standby mode when the disc stops rotating and other electronic devices shut down. If the disk is in sleep mode, the energy consumption is zero.

Fig.1 The energy consumption modes of disk

For most of the hard disk products, if there is no I/O request to disk for a period of time, the disk will switch from idle to standby mode in order to reduce the energy consumption of rotation. For one disk, within a period of observation time Td, ΔErotateis the energy saved by mode switching.

On the other hand, when processing I/O requests, the seek motor drives the magnetic head to seek the specified track. One seek movement spends several seconds and consumes considerable electric energy that is usually more than twice as regular rotation. Moreover, when the disk processes random I/O, the magnetic head has to seek back and forth, which leads to higher energy consumption. For one disk, in period Td, the energy consumption caused by magnetic head is

(2)

Among them, Pseekand Protateare the electric power consumed for rotating and seeking respectively, nseekis the total number of seek movements in Δtiwhich is the time consumption for the i-th seek movement, Protate·Δtigives the energy consumption of rotating idly while seeking.

2.2 Energy model of streaming media storage system

According to the I/O characteristics of streaming media, as well as the requirements on system scalability, at present, most of the streaming media storage systems in the industry are deployed in horizontal striping model on logical volume (namely LUN, Logical Unit Number) distribution in disk group (DG). The horizontal striping model is illustrated in Fig.2. Besides, due to the isolation of performance between different DGs, without loss of generality, this paper only discusses the situation in a single DG in streaming media storage system.

Fig.2 The horizontal striping model

Assume that the number of disks in this DG is ND, the number of LUNs is NL. For the entire DG in horizontal striping, in period Td, the energy saved by mode switching is

(3)

As is shown in Fig.2, the magnetic heads of NDdisks in horizontal striping will be driven together when processing random I/O, which further results in considerable cumulative currents and great increase in energy consumption. Assuming that the number of seeks between different LUNs is Nskipin Td, the average number of seeks between different tracks in the same LUN is Ntrack, the seeking energy consumption given by magnetic heads seeking of this DG is

(4)

In Eq.(4), Nseekis the total seek times in period Td, namely Nseek=Nskip×Ntrack. For streaming media accessing, the addresses of the I/O requests reaching together are usually contiguous and the I/O blocks are quite large, therefore Nskip>>Ntrack. Furthermore, ideally it can be assumed that Nseek≈Nskip. Δtiis the time consumption for the i-th seek, and Δtiis usually about several milliseconds. Pseekand Protateare the electric power consumed for disks rotating and seeking respectively, and Pseekis greater than Protateby several dozens of times generally. Hence, for some specific application scenarios such as streaming media, it is reasonable to take the I/O patterns and characteristics of the application into account and to further work out some scheduling schemes to ease the frequent seeking movements between LUNs in order to reduce energy consumption.

(5)

3 Energy-saving scheduling scheme based on I/O stream

3.1 Overview

As is mentioned before, the optimized I/O requests from different initiators will be mixed disorderly when they are reaching the storage system concurrently, which leads to increasing energy consumption.

Fig.3 The I/O requests accessing the disk group

For the traditional and general storage systems, the mixed I/O requests are simply handled by the FCFS scheme. Therefore, the mixed I/O requests in these storage systems usually result in rapid growth of energy consumption and great decrease of I/O throughput performance.

This paper proposes an energy-saving scheduling scheme based on I/O stream (ES-IOS). ES-IOS can take the advantage of the I/O characteristics of streaming media and reorganize the mixed and disordered I/O requests into “streams”. Then these streams can be treated as the basic units in scheduling. Technically, ES-IOS includes two main points, a priority-based weighted stream scheduling algorithm (PWSS) and a regression-fitting-based popularity prediction algorithm (RFPP).

The ES-IOS scheme can reduce energy consumption and improve the throughput performance of streaming media storage systems via limiting the frequent seeks between LUNs. As the cost, it may cause some extra access delay.

3.2 Priority-based weighted stream scheduling algorithm (PWSS)

Some studies[22,23]pointed out that the popularity and the I/O requests of streaming media obey Zipf distribution. That is, if there are M streaming media files in the storage system and the files have been sorted according to popularity in descending order, the access

Fig.4 The energy-saving scheduling scheme based on I/O stream

probability of the i-th file will be C/i1-θ. The parameter C is

C=1/(1+1/2+1/3+…+1/M)

(6)

That is the Zipf distribution, also called Chervenak law. In Eq.(7), θ is a certain slope index. If θ=0, it is a standard Zipf distribution. Z(i) means the access probability of the i-th file.

(7)

The Zipf distribution indicates the I/O characteristics of streaming media that some hot programs or media files will attract most of the accessing requests in a period. In PWSS, the popularity of each I/O stream is equivalent to the sum of popularity of all the media files on the LUN.

Therefore, PWSS assigns a priority weight Wito each I/O stream. Wiis determined by the popularity factor popiof the stream Si, namely Wi=α·popi, α is the conversion ratio. According to the parameter Wi, PWSS assigns more continuous service time to the hot streams and reversely less service time to the cold ones.

Based on the I/O access records in the last time period, RFPP can predict the popularity factor of each stream. The detail of the RFPP will be introduced later. In order to ensure the fairness in stream scale, as well as limiting the access delay of cold streams, PWSS sets up a Wake up Timer (WT) of each stream when they are processing I/O requests. When there is a WT timeout of Si, Siwill be added into the time-out queue with the highest priority.

(8)

The procedure of PWSS algorithm is as follows:

Step 1: Initialization. All of the NLstreams will be initialized by mapping them to the corresponding LUNs on the DG. The priority weight Wiof each stream should be calculated from the popularity factor popi. An Update Timer (UT) will be setup for renewing the weight values of the I/O streams periodically.

Step 3: Polling and serving. Polling and serving is the key step of PWSS which is illustrated by flow chart in Fig.5. Each stream is served for an occupy time by polling. When polling to a stream, PWSS first checks the time-out queue Qtimeout. If there are I/O streams to be processed in the Qtimeout, these streams must be served with the highest priority; If there is no I/O stream in the Qtimeout, it will continue to serve the current stream.

Fig.5 The procedure of polling and serving in PWSS

Step 4: Updating weights. When there is a UT timeout, the popularity factor of each stream should be calculated via the prediction algorithm RFPP. Then the priority weight will be obtained from the popularity factor. To be mentioned, because the stream being served is locked up by PWSS, the weights of the I/O streams except for the stream being served can be updated. Once the I/O stream has been processed and unlocked, PWSS will update the weight of this stream immediately.

3.3 Regression-fitting-based popularity prediction algorithm (RFPP)

Priority weight Wiin PWSS is determined by popularity factor popiof stream Siwhich is the total access probability of LUNi. Therefore, popularity factor popiis defined to describe the overall popularity of the streaming media files in LUNi. In other words, popiequals to the sum of the access probabilities of the files in LUNi. Because of the time-varying characteristics, hardly can the popularity be obtained exactly at some moment in the future. Therefore, this paper designs a regression-fitting-based popularity prediction algorithm (RFPP) that can predict the popularity factor of each I/O stream according to the I/O access records over a time period. The RFPP can obtain high prediction accuracy and low complexity. Therefore, RFPP is suitable for real-time dynamic prediction.

The main idea of RFPP shown as follows

Total N·K I/O requests counted in the K periods do not only indicate the access probability of every LUN, but also describe the changing trend of popularity. According to the statistical results in the K periods, RFPP algorithm can predict the popularity of every LUN in the (K+1)-th period via a regression fitting method.

By considering the prediction accuracy and the system memory usage in regression fitting the popularity of streaming media files, Yang, et al.[25]pointed out that the number of periods K should be set to 5～10. Because the popularity will not fluctuate frequently in this range, quadratic regression can meet the demand of accuracy. Furthermore, since the matrix multiplication in the regression fitting method has the time complexity of O(n3), quadratic regression can lead to less operation time than cubic, quartic or some higher orders.

(9)

Assuming that k2=k2, k1=k, the quadratic regression model can be converted into a binary linear regression model.

(10)

Assuming that T is the coefficient matrix of Eq.(10).

(11)

(12)

Then the popularity factor of LUNiin the (K+1)-th

Fig.6 The procedure of RFPP algorithm

4 Experimental evaluation

A storage system for streaming media is deployed in this work. In order to simplify the problem, the storage system has a single disk group that is divided into several LUNs in horizontal striping. A number of streaming media files are stored in this storage system. In a realistic application environment, this storage system has been utilized to provide streaming media services for consumer electronics terminals. Over a period of time, the I/O requests from the actual streaming media users to each LUN are monitored and recorded. In order to facilitate further analysis, without loss of generality, two groups of access records have been selected, named as Trace1 and Trace2 respectively.

In Fig.7, the continuous 24 hours of streaming media access records in the storage system are demonstrated. The horizontal axis presents the sampling moments. The vertical axis describes the number of I/O requests accessing the LUNs to describe the popularity of each stream.

As can be seen from Trace1, the popularities vary greatly for each LUN. LUN1 is the hottest, followed by LUN2, and the popularity of LUN6 is the lowest. In addition, the popularities of the LUNs keep relatively stable without obvious time variability. For Trace2, the

(a)

(b)

popularities of LUN1, LUN3, and LUN5 have varied greatly over the observation time, which indicates the time-varying characteristics of streaming media popularity.

4.1 Experiment setup

The simulation experiment is based on the cases of Trace1 and Trace2. The energy saving, I/O throughput, delay and jitter performance are evaluated in this section. The general FCFS scheduling algorithm is acting as comparison.

The disk group in the experiment consists of ten data disks and six LUNs in horizontal striping. Referring to the general specifications of the disks at present, the simulation parameters of the disk are set as Table 1.

Table 1 Disk parameters

The polling weight of each stream in PWSS is determined by the popularity factor, namely Wi=α·popi. α is the conversion ratio. Assuming that it takes Tpollto poll each one of the I/O streams, the I/O arrival rate of stream Siis λi, the size of each I/O is Block, and the I/O processing rate of the storage system is μ MB/s.

(13)

The simulation parameters of the experiment are illustrated in Table 2. In order to simplify the situation without losing generality, the I/O requests accessing the same LUN can be reorganized as an I/O stream. Then the conversion ratio can be estimated as α≥0.1. Since α is the ratio of the serving capacity of storage system and the I/O load, α should be set to α=0.12.

Table 2 Simulation parameters

4.2 Experiment results

4.2.1 Energy consumption

The energy consumption is the primary point that this paper focuses on. Fig.8 shows the comparison of energy consumption for Trace1 and Trace2 when implementing PWSS algorithm and FCFS algorithm. The PWSS algorithm has a significant energy saving effect. Compared with FCFS, PWSS can approximately deliver 38.3% and 39.2% energy-saving ratio on average in Trace1 and Trace2 respectively.

(a)

(b)

The reason is that the ES-IOS scheme can take the advantage of the I/O characteristics of streaming media and reorganize the mixed and disordered I/O requests into streams. Furthermore, according to the media popularity, PWSS assigns more continuous service time to the hot streams and reversely less service time to the cold ones. Therefore, PWSS can limit the frequent seeks between different streams and reduce the energy consumption significantly.

4.2.2 I/O throughput

When implementing PWSS, the improvement of I/O throughput is coming together with the energy saving. The experimental results in Fig.9 have proved it. Compared with FCFS, PWSS can obtain 28.9% and 27.8% improvement of I/O throughput on average in Trace1 and Trace2 respectively.

4.2.3 Delay and delay jitter

As is analyzed earlier, the negative effect of PWSS is some extra access delay. Since the ES-IOS scheme divides the mixed I/O requests into I/O streams and the streams are the basic units in scheduling, it is proper to evaluate the delay in stream scale.

(a)

(b)

As is shown in Fig.10, PWSS has brought 29.70% and 27.91% more delay than FCFS in stream scale for Trace1 and Trace2 respectively. However, because of

(a)

(b)

the weighted stream scheduling, the extra delay of PWSS is smoother than FCFS. Therefore, the extra smooth delay can be relieved effectively by some cache schemes, e. g. DRAM, SSD, or even redundant disks.

Real-time is one of the key features of streaming media services. In terms of QoS, streaming media put strict requirements on the delay and delay jitter[26], e.g. audio stream allows a maximum delay of 250ms and delay jitter of 10ms[27]. The delay jitter (jitter for short) is defined as the variability of the delay, which describes the playing fluency of streaming media.

J=|d-E(d)|

(14)

In Eq.(14), d is the delay, J is the jitter, and E(d) is the average delay of the streaming media service. The jitter can affect the QoS of media streaming greatly. Tulu et al.[26]presented that the streaming media users usually can tolerate some delays, but can hardly have patience with jitters.

In Fig.11, the delay jitters in stream scale of PWSS are 68.89% and 66.23% less than FCFS in Trace1 and Trace2 respectively. Hence, despite causing extra delay, PWSS has smoothed the delay jitter considerably and may further improve the QoS of the streaming media service.

(a)

(b)

5 Conclusion

In this work, an energy-saving scheduling scheme based on I/O stream (ES-IOS) has been presented. The ES-IOS scheme can take the advantage of the I/O characteristics of streaming media and reorganize the mixed and disordered I/O requests into “streams”. Technically, ES-IOS includes two main points as a priority-based weighted stream scheduling algorithm (PWSS) and a regression-fitting-based popularity prediction algorithm (RFPP). PWSS can schedule the I/O streams in weighted queue based on the priority to limit energy consumption. The priority of each stream is determined by its popularity. According to the I/O access records over a time period, RFPP can predict the popularity of each stream via regression fitting. Based on the popularities, PWSS assigns more continuous service time to the hot streams and reversely less service time to the cold ones. ES-IOS scheme can reduce the energy consumption and improve the throughput performance of streaming media storage systems via limiting the frequent seeks between LUNs. As for the cost, ES-IOS may cause some extra access delay. However, the delay jitter in stream scale can be smoothed by ES-IOS considerably.

According to the statements in Section II, the ES-IOS scheme is based on exploiting the I/O characteristics of streaming media. For the future directions, it is reasonable to improve energy efficiency and I/O performance via adjusting the data distribution of the streaming media storage systems, for instance migrating data based on popularity. Then, these two aspects should be combined to achieve more complete solutions.

[ 1] Zakerinasab M R, Wang M. A cloud-assisted energy-efficient video streaming system for smartphones. In: Proceedings of the IEEE/ACM 21st International Symposium on Quality of Service, Montreal, Canada, 2013. 1-10

[ 2] Yang L H, Zhou J, Gong W H, et al. Energy-efficient replacement schemes for heterogeneous drive. Journal of Computer Research and Development, 2013, 50(1): 19-36

[ 3] Gurumurthi S, Sivasubramaniam A. Thermal issues in disk drive design: Challenges and possible solutions. ACM Transactions on Storage (TOS), 2006, 2(1): 41-73

[ 4] Tian L, Dan F, Yue Y L, et al. Survy on power-saving technologies for disk-based storage systems. Computer Science, 2010, 37(9): 1-5

[ 5] Gurumurthi S, Sivasubramaniam A, Kandemir M, et al. DRPM: Dynamic speed control for power management in server class disks. In: Proceedings of the IEEE 30th Annual International Symposium on Computer Architecture, 2003. 169-179

[ 6] Huang H, Hung W, Shin K G. FS2: Dynamic data replication in free disk space for improving disk performance and energy consumption. In: Proceedings of the 20th ACM symposium on Operating systems principles, New York, USA, 2005. 263-276

[ 7] Papathanasiou A E, Scott M L. Energy efficient prefetching and caching. In: Proceedings of the 2004 USENIX Annual Technical Conference, Berkeley, USA, 2004. 255-268

[ 8] Weddle C, Oldham M, Qian J, et al. PARAID: A gear-shifting power-aware RAID. ACM Transactions on Storage (TOS), 2007, 3(3): 13

[ 9] Colarelli D, Grunwald D. Massive arrays of idle disks for storage archives. In: Proceedings of the 2002 ACM/IEEE Conference on Supercomputing, Los Alamitos, USA, 2002. 1-11

[10] Pinheiro E, Bianchini R. Energy conservation techniques for disk array-based servers. In: Proceedings of the 18th ACM Annual International Conference on Supercomputing, New York, USA, 2004. 68-78

[11] Li D, Wang J. eRAID: A queueing model based energy saving policy. In: Proceedings of 14th IEEE International Symposium on Modeling, Analysis, and Simulation of Computer and Telecommunication Systems, 2006. 77-86

[12] Zhu Q, Zhou Y. Power-aware storage cache management. IEEE Transactions on Computers, 2005, 54(5): 587-602

[13] Zhu Q, Chen Z, Tan L, et al. Hibernator: Helping disk arrays sleep through the winter. In: Proceedings of the 20th ACM Symposium on Operating systems principles (SOSP), New York, USA, 2005. 177-190

[14] Narayanan D, Donnelly A, Rowstron A. Write off-loading: Practical power management for enterprise storage. ACM Transactions on Storage (TOS), 2008, 4(3): 10

[15] Pinheiro E, Bianchini R, Dubnicki C. Exploiting redundancy to conserve energy in storage systems. In: Proceedings of the Joint International Conference on Measurement and Modeling of Computer Systems, New York, USA, 2006. 15-26

[16] Useche L, Guerra J, Bhadkamkar M, et al. EXCES: External caching in energy saving storage systems. In: Proceedings of the IEEE 14th International Symposium on High Performance Computer Architecture, Salt Lake City, USA, 2008. 89-100

[17] Shim H, Kim J S, Maeng S. BEST: Best-effort energy saving techniques for NAND flash-based hybrid storage. IEEE Transactions on Consumer Electronics, 2012, 58(3): 841-848

[18] Kim J, Yang A, Song M. Exploiting flash memory for reducing disk power consumption in portable media players. IEEE Transactions on Consumer Electronics, 2009, 55(4): 1997-2004

[19] Payer H, Sanvido M A, Bandic Z Z, et al. Combo drive: Optimizing cost and performance in a heterogeneous storage device. In: Proceedings of the 1st Workshop on Integrating Solid-state Memory into the Storage Hierarchy, 2009, 1(1): 1-8

[20] Chen F, Koufaty D A, Zhang X. Hystor: making the best use of solid state drives in high performance storage systems. In: Proceedings of the ACM International Conference on Supercomputing, Seattle, USA, 2011. 22-32

[21] Ryu W, Song M. Design and implementation of a disk energy saving scheme for media players which use hybrid disks. IEEE Transactions on Consumer Electronics, 2010, 56(4): 2382-2386

[22] Chesire M, Wolman A, Voelker G M, et al. Measurement and analysis of a streaming media workload. In: Proceedings of the 3rd USENIX Symposium on Internet Technologies and Systems, San Francisco, USA, 2001. 1-12

[23] Jin S, Bestavros A, Iyengar A. Accelerating Internet streaming media delivery using network-aware partial caching. In: Proceedings of the IEEE 22nd International Conference on Distributed Computing Systems, Vienna, Austria, 2002. 153-160

[24] Luo Z G, Sun W, Wang X G. A transferring cost based cache replacement algorithm for streaming media and its performance evaluation. Journal of China Institute of Communications, 2004, 25(2): 61-67

[25] Yang C D, Yu Z W, Wang X G, et al. Proxy cache replacement algorithm based on popularity prediction of streaming media file. Computer Engineering, 2007, 33(7): 99-129

[26] Tulu B, Chatterjee S. Internet-based telemedicine: An empirical investigation of objective and subjective video quality. Decision Support Systems, 2008,45(4): 681-696

[27] Qiu H, Li Y F, Wu J X. Analysis on network delay for temporal structure guarantee of streaming media. Journal of Electronics & Information Technology, 2009, 31(10): 2287-2293

Shang Qiuli, born in 1987. He is currently a Ph.D. candidate at Institute of Acoustics, Chinese Academy of Sciences. He received his B.S. degree in Information Engineering from Beijing University of Posts and Telecommunications, China, in 2011. His current research interests include streaming media, storage system and energy-saving technology.

10.3772/j.issn.1006-6748.2015.03.016

①Supported by the National High Technology Research and Development Programme of China (No. 2011AA01A102).

②To whom correspondence should be addressed. E-mail: shangql@dsp.ac.cn Received on Feb. 27， 2014, Zhang Wu, Guo Xiuyan, Chen Xiao, Ni Hong