You are on page 1of 7

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220909571

A Power Consumption Model of a Storage Server

Conference Paper · September 2011


DOI: 10.1109/NBiS.2011.64 · Source: DBLP

CITATIONS READS
17 572

4 authors:

Takuro Inoue Ailixier Aikebaier

11 PUBLICATIONS   113 CITATIONS   
National Institute of Information and Communications Technology
130 PUBLICATIONS   1,663 CITATIONS   
SEE PROFILE
SEE PROFILE

Tomoya Enokido Makoto Takizawa


Rissho University Hosei University
408 PUBLICATIONS   3,695 CITATIONS    1,117 PUBLICATIONS   7,800 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Energy-efficient distributed systems View project

All content following this page was uploaded by Makoto Takizawa on 15 March 2019.

The user has requested enhancement of the downloaded file.


2011 International Conference on Complex, Intelligent, and Software Intensive Systems

A Power Consumption Model for Storage-based Applications

Takuro Inoue∗ , Makoto Ikeda∗ , Tomoya Enokido† , Ailixier Aikebaier∗ and Makoto Takizawa∗
∗ Seikei University, Tokyo, Japan
Email: takuro.burton@gmail.com
makoto.ikd@gmail.com
alisher.akber@computer.org
makoto.takizawa@computer.org
† Risho University, Tokyo, Japan
Email: eno@ris.ac.jp

Abstract—In order to realize digital ecosystems, the total Files in storage drives like hard disk drives are mainly
electric power consumption of servers to perform applications manipulated in the storage-based applications. Here, the
has to be reduced in information systems. Applications are electric power of a computer is consumed to access to the
classified into computation, communication, and storage based
types. In the computation and communication based appli- storage drives in addition to CPU resources. In this paper,
cations, CPU and communication resources of computers are we discuss the power consumption model of a computer
mainly consumed, respectively. In the storage-based applica- to perform the storage-based applications. We first measure
tions, the electric power is consumed to access to storage the power consumption of a computer where files in types
drives like hard disk drives (HDDs) and solid state drives of storage drives, HDD (hard disk drive) and SSD (solid
(SSDs). In this paper, we measure how much electric power
of a server computer is consumed to read and write data in state drive) are manipulated by applications processes. Then,
types of storage drives. Then, we discuss a power consumption by abstracting essential properties from the experimental
model of a computer to perform a storage-based applications results, we obtain a power consumption model to read and
based on the measurement of the power consumption of write a file in the types of storage drives where the total
storage drives. Here, the power consumption rate of a server power consumption is proportional to the size of a file. We
is maximum if at least one process is performed. Otherwise,
the power consumptionrateis minimum. This is a simple power do not consider the power consumption of each component,
consumption model. i.e. storage drive of a computer at a micro level. We rather
I. I NTRODUCTION consider the total power consumption of the whole computer
Digital ecosystems [1] have to be realized in order to to perform application processes at a macro level. In the
reduce the consumption of natural resources, especially oil paper, we show a simple power consumption model [3] for
in the Earth. In information systems, we have to reduce the storage-based applications.
total electric power consumption of server to perform various In section II, we present the experimental results on the
types of applications. Applications in information systems power consumption of a computer in storage-based appli-
are classified into three types; computation, communication, cations. In section III, we discuss the power consumption
and storage-based types of applications. In the computation- model obtained by abstracting dominating properties from
based applications [2], CPU resources are mainly consumed. the experimental results.
The single-read and multi-read power consumption models
II. E XPERIMENTAL R ESULTS
are proposed for computation-based applications [3][4][5].
In the communication-based applications, a large volume A. Measurement of power consumption
of data are transmitted in networks likes file transfer ap- In order to model the power consumption of a computer
plications like FTP [8]. Algorithms to select a server in in storage-based applications, we first measure the power
a server set are discussed so that not only the deadline consumption of a computer to access to secondary storage
constraint holds but also the power consumption can be drives. We do not measure the power consumption of each
reduced. The power consumption of an FTP sever depends storage drive component at a micro level of a computer like
on the total power transmission rate and number of clients hard disk drives, memory, and fans. It is very difficult to
to which files are concurrently transmitted [2]. Algorithms make a model showing the power consumption to perform
for selecting an FTP sever in a set of servers are discussed application programs on a computer by synthesizing the
so that not only the transmission time but also the power power consumption of each component. We rather mea-
consumption can be reduced. The power consumption model sure the power consumption of a whole computer where
of a computer to perform a general type of application application processes are performed at a macro level. We
where both computation and communication resources are measure the electric power consumption of the computer by
consumed is discussed [5]. using the power meter Watts Up?.net [7]. Here, the power

978-0-7695-4373-4/11 $26.00 © 2011 IEEE 612


DOI 10.1109/CISIS.2011.101
consumption of the computer can be measured every one 100
HDD
second. SSD

Execution time TR [sec]


80
We consider a couple of types of the secondary storage
drives; hard disk drive (HDD) and solid state drive (SSD) in 60
a personal computer (PC) as shown in Table 1. Processes
40
which read and write files in HDD and SSD types of
storage drives are executed on Linux operating system (Cent 20
OS) in the computer. Here, no other application process is
0 0
performed on the computer. We measure the electric power 10 10
1 2
10
3
10 10
4

File size s [MB]


Table I
PC AND DRIVES Figure 1. Execution time for read.
15000

Power consumption ER [Ws]


PC@ CPU : Intel(R) Core(TM) i7 HDD
Memory : 6.00GB DDR3@3 x 2048
SSD
OS : Cent OS 64 bit
HDD Capacity : 500GB 10000
Rotation : 7200 rpm
Read speed : 121.45MB/s
Write speed : 106.27MB/s
Interface : Serial ATA300 5000
SSD Capasity : 60GB
Read Speed : 197.02MB/s
Write Speed : 93.89MB/s
Interface : Serial ATA300 0 0 1 2 3 4
10 10 10 10 10
File size s [MB]
consumption of the computer by using the power meter
Watts Up?.net [7]. Here, the power consumption of the Figure 2. Total power consumption for read.
computer can be measured every one second. 112
Power consumption rate er [W]

HDD
SSD
B. Reading files 111

First, We consider a process which just reads a file of 110


size s [MB(yte)] by using a read system call in a type
τ (∈ {H(HDD), S(SSD)}) of storage drive. Data size 109

of 1024 [B] are read into a buffer in one read system call. 108
Figures 1 and 2 show the execution time [sec] and total
power consumption [Wsec] of the computer to read a file of 107
0 2000 4000 6000 8000 10000
size s [MB], respectively. For example, it takes 91.6 [sec] File size s [MB]
and 51.2 [sec] to read a file of 1 [GB] in HDD and SSD types
Figure 3. Power consumption rate for read.
of drives, where the electric power 1001.2 [Wsec] and 549.8
[Wsec] are totally consumed in HDD and SSD, respectively.
The execution time [sec] to read a file is proportional to the
erH (s) = 111 [W/sec] and for s > 2000 [MB]. In the SSD
size s of the file in HDD and SSD types of drives as follows
drive, erS (s) = 0.0006s + 107.19 [W/sec] for s ≤ 2000
:
[MB] and erS (s) = 109 [W/sec] for s > 2000 [MB].
T RH (s) = 0.0084s. Next, multiple processes p1 , ... , pm reading files are
T RS (s) = 0.0051s. concurrently performed for a storage drive type τ (∈
{H(HDD), S(SSD)}). We consider the following two
The total power consumption [Wsec] to read a file is also cases:
proportional to the file size s as : 1) Single-read : every process pi reads the same file f in
ERH (s) = 0.940s. a type τ of storage drive.
2) Multi-read : every pair of processes pi and pj read
ERS (s) = 0.562s. different files fi and fj (fi = fj ) in a type τ of drive,
Figure 3 shows the power consumption rate erτ (s) respectively.
[W/sec] to read a file of size s in a type τ of storage We assume the size of every file fi is the same s as the file
drive. The power consumption rate [W/sec] is considered f (i = 1, · · · , m). Figures 4, 5, and 6 show the execution
to be constant for file size s ≥ 2000 [MB]. In the HDD time, total power consumption, and power consumption rate
drive, erH (s) = 0.0018s + 107.35 for s ≤ 2000 [MB] and for of the computer m concurrent processes to read files of

613
100 Figure 6 shows the power consumption rate [W/sec]
HDD
SSD erτ2 (s) for two processes to read files of size s in a type τ of
Execution time TR [sec]

80
storage drive. Here, the power consumption rate is constant
60 for s ≥ 1000 [MB]. Here, the power consumption rates
2
erH (s) = 0.0016s+107.4 and erS2 (s) = 0.0007s+107.0 for
40 2
s ≤ 1000 [MB]. erH (s) = 111 [W/sec] and erS2 (s) = 109
20 [W/sec] for s > 1000 [MB]. Compared with Figure 3,
the power consumption rate for m = 2 is the same as
0 0
10 10
1 2
10
3
10
4
10 m = 1. That is, the power consumption rate is constant
File size s [MB] independently of m.
Figure 4. Execution time for read (m = 2).
The average execution time and average power con-
sumption of each process are increased if more number
15000
m of processes are concurrently performed. The average
Power consumption ER [Ws]

HDD
SSD execution time AT Rτm (s) of each process to read a file of
10000 size s concurrently with (m − 1) processes is T Rτm (s)/m.
AERτm (s) shows the average power consumption of each
process, which is ERτm (s)/m. Table II shows AT Rτm (s)
5000 and AERτm (s) for a storage drive type τ ∈ {H, S } where
s shows the size [MB] of each file. According to Table
II, the average execution time AT Rτ, (s) of each process
0 0
10 10
1
10
2
10
3
10
4
is independent of m.
File size s [MB]
Table II
Figure 5. Total power consumption for read (m = 2). AVERAGE EXECUTION TIME AND POWER CONSUMPTION ( READ )
112 ATR [sec] AER [sec]
Power consumption rate er [W]

HDD
SSD m H S H S
1 0.0084 s 0.0051 s 0.94 s 0.562 s
110 2 0.00865 s 0.0051 s 0.957 s 0.557 s
3 0.853 s 0.517 s 0.944 s 0.560 s
4 0.0855 s 0.00515 s 0.978 s 0.567 s
s = size [MB] of a file.
108

106 C. Writing files


0 2000 4000 6000 8000 10000
File size s [MB]
We measure the execution time and total power consump-
Figure 6. Power consumption rate for read (m = 2). tion to write files in HDD and SSD types of storage drives.
A process writes data of size s [MB] to a file f. One unit
data of sizes 1024 [MB] is written into a file f by using a
size s in HDD and SSD for m = 2. T Rτm (s) and ERτm (s) write system call. Figures 7 and 8 show the execution time
show the total execution time and power consumption for and power consumption to write data of size s, respectively,
each of m concurrent processes to read a file of size s [MB] in HDD and SSD. The execution time to write a file is
in a drive type τ , respectively. The total execution time of proportional to the size s of the file in HDD and SSD as
m processes to read files of size s [MB] is given for m = 2 reading a file. For example, it takes 0.24 and 0.21 [sec] to
as follows : write a file f of 100 [MB] to HDD and SSD, respectively.
2
T RH (s) = 0.0173s. Here, SSD is 1.14 times faster than HDD. It takes 3.02 [sec]
and 2.46 [sec] for a file of 1 [GB]. By using the least-squares
T RS2 (s) = 0.0102s. method, we obtain the execution time [sec] with respect to
The total power consumption of the computer to read a file size s [MB] of file for H(HDD) and S(SSD) types of drives
of size s is given as follows : is given as follows :
2
ERH (s) = 0.992s. T WH (s) = 0.0072s.
ERS2 (s) = 0.603s. T WS (s) = 0.0053s.
The execution time T Rτ2 (s) and total power consumption The total power consumption [Wsec] with respect to size s
ERτ2 (s) are proportional to the file size s for τ ∈ {H, S}. [MB] of a file for H(HDD) and S(SSD) types of drives is

614
given as follows : 80
HDD
SSD

Execution time TW [sec]


EWH (s) = 0.8676s. 60
EWS (s) = 0.6445s.
40
For example, the power consumption 55.16 [Wsec] and
442.04 [Wsec] are consumed to write a file f of size 100
20
[MB] and 1 [GB] in SSD, respectively. On the other hand,
96.58 [Wsec] and 1001.22 [Wsec] are consumed to write a 0 0 1 2 3 4
file f of size 100 [MB] and 1 [GB] to HDD, respectively. 10 10 10 10 10
File size s [MB]
The power consumption is 0.57 times and 0.54 times less
consumed in SSD than HDD, to write files of 100[MB] and Figure 7. Execution time for write.
1[GB], respectively.

Power consumption EW [Ws]


10000
Figure 9 indicates the power consumption rate [W/sec]. HDD
The power consumption rate [W/rec] is constant for file size 8000 SSD
s (≥ 2000 [MB]). Here, ewH (s) = 0.012s + 110.42 and
6000
ewS (s) = 0.0141s + 112.67 for s ≤ 2000 [MB]. ewH (s) =
120 [W] and ewS (s) = 120 [W] for s ≥ 2000 [MB]. 4000
Next, multiple processes p1 , ... , pm (m ≥ 1) concurrently
2000
write data into files f1 , ... , fm in a type τ of storage drive,
respectively. Here, each pair of files fi and fj are different, 0 0 1 2 3 4
fi = fj . Each process pi writes data of the same size s 10 10 10 10 10
File size s [MB]
[MB] into a file fi . We measure the total execution time
T Wτm (s) and total power consumption EWτm (s) for each Figure 8. Total power consumption for write.
of m concurrent processes to write data of size s to a file 140
in a type τ of storage drive. Figures 10 and 11 show the
Power consumption rate ew [W]

HDD
SSD
execution time T Wτ2 (s) and the total power consumption 130
EWτ2 (s) for file size s, respectively, which are proportional
to the file size s : 120
2
T WH (s) = 0.0182s.
T WS2 (s) =
110
0.0136s.
2
EWH (s) = 2.12s. 100
2 0 2000 4000 6000 8000 10000
EWS (s) = 1.59s. File size s [MB]

2
Figure 12 shows the power consumption rate ewH (s) = Figure 9. Power consumption rate for write.
2
3.44s − 1147.2 and ewS (s) = 2.19s − 710.3 for s ≤ 2000
2
[MB]. The power consumption rate ewH = 120 [W/sec] Table III
2 AVERAGE EXECUTION TIME AND POWER CONSUMPTION ( WRITE )
and ewS = 120 [W/sec] for s > 2000 [MB]. Compared
m
with Figure 9, the power consumptions ewH and ewSm are ATW [sec] AEW [sec]
1 m H S H S
constant independently of m for s > 2000 [MB], ewH = 1 0.0072 s 0.0053 s 0.868 s 0.645 s
2 1 2
ewH and ewH = ewS . 2 0.091 s 0.0068 s 1.058 s 0.795 s
The average execution time AT Wτm (s) of each process to 3 0.0101 s 0.00713 s 1.161 s 0.837 s
4 0.0103 s 0.00723 s 1.196 s 0.0848 s
write a file of size s concurrently with (m − 1) processes is s = size [MB] of a file.
T Wτm (s)/m. The average power consumption AEWτm (s)
of each process is EWτm (s)/m. Table III shows the average
execution time AT Wτm (s) and the average power consump-
III. P OWER C ONSUMPTION M ODEL
tion AEWτm (s).
m
Figure 13 shows the average execution time AT WH (s) = Based on the experimental results, we discuss a power
m
T WH (s)/m of each process for file size s and the number consumption model for storage-based applications. Ac-
m of concurrent processes. Here, s = 4 [GB] for s = 4 cording to the experimentations presented in the preced-
[GB]. By the least-squares method, AT WH m
(s) = 0.001m+ ing subsection, the average execution time AT Rτm (s) and
0.0066. Figure 14 shows the average power consumption AT Wτm (s) to read and write a file of size s in a type τ
AT Wsm (s) for the number m of concurrent processes. By of storage drive is proportional to the size s of the file and
the least-squares method, AT WSm (s) = 0.0006m + 0.0051. depends on the number m of concurrent processes. From

615
250

Average execution time ATWH [sec]


HDD 45
SSD ATW
Execution time TW [sec]

m
200
40
150

100 35

50
30
0 0 1 2 3 4
10 10 10 10 10
File size s [MB] 25
1 2 3 4
m (s = 4GB)
Figure 10. Execution time for write (m = 2).
4
x 10 Figure 13. Average execution time for HDD .
3
HDD
Power consumption EW [Ws]

Average execution time ATWS [sec]


2.5 SSD 30
ATWS

m
2
28
1.5
26
1

0.5 24
0 0 1 2 3 4 22
10 10 10 10 10
File size s [MB]
20
Figure 11. Total power consumption for write (m = 2). 1 2 3 4
m (s = 4GB)
160
Power consumption rate ew [W]

HDD
SSD Figure 14. Average execution time for SSD .
150

140

130 processes are concurrently performed, the longer it takes


and the more power consumption is consumed to perform
the processes. Table IV shows the coefficient ατt (m) for
120

110 τ ∈ {H, S} and t ∈ {R, W }.


100 Nj (t) is the number of processes concurrently performed
0 2000 4000 6000
File size s [MB]
8000 10000
on a computer cj at time t. The power consumption rates
erjτ (t) and ewjτ (t) [W/sec] to read and write a file,
Figure 12. Power consumption rate for write (m = 2). respectively, is given for a type τ of storage drive in a
computer cj at time t.

the experimental results, we obtain the following formulas maxEj if Nj (τ ) ≥ 1.
erjτ (t)
AT Rτm (s) and AT Wτm (s) to show the average execution minEj otherwise
time AT Rτm (s) and AT Wτm (s) [sec] of each process to If at least one process is performed, maxEj is consumed
read and write a file of size s [KB] in a type τ ({S, H}), at a computer Cj . minEj shows the minimum power
respectively, of storage drive concurrently with (m − 1) consumption rate of a computer cj when no process is
processes; performed. If at least one process to read and write a file is
T Rτm (s) = ατR (m) · AT Rτ1 (s). (1) performed on a computer cj , the maximum power maxEj is
consumed. If no process is performed, the minimum power
T Wτm (s) = ατW (m) · AT Wτ1 (s). (2) minEj is consumed. This is similar to the simple power
ατt (m) shows the degradation factor. The more number of consumption model for computation-based applications [6]
Here, ατR (m) is independent of m.
Table IV
αtτ (m) IV. C ONCLUDING R EMARKS
In this paper, we discussed the power consumption model
t H S
R 1 1 for storage-based applications. First, we measured the real
W 1.001 m 1.0006 m power consumption of a whole computer with two types of
m = number of concurrent processes. storage drives, HDD and SSD. The execution time to read

616
and write a file is proportional to the size of the file. Based
on the experimental results, we made the simple power
consumption model. The maximum power where a server
consumes the maximum power if at least one process is
performed. By using the model, we would like to discuss
how to select a storage server in a set of the storage servers
so that the total power consumption can be reduced.
ACKNOWLEDGMENT
This research is partially supported by the strategy re-
search of Seikei University and MEXT, Grant-in-Aid for
Building Strategy Research Infrastructure.
R EFERENCES
[1] H. Dong, F .K. Hussain, and E. Chang : Semantic Service
Retrieval and QoS Measurement in the Digital Ecosystem
Environment. Proc. of the Intelligent and Software Intensive
Systems (CISIS), page:153-160, 2010 .

[2] T. Enokido, A. Aikebaier, S. Misbah Deen, and M. Takizawa


: Power Consumption-based Server Selection Algorithms for
Communication-based Systems. Proc. of the 13th Interna-
tional Conference on Network-based Information Systems
(NBiS2010), page:201-208, 2010.

[3] T. Enokido, A. Aikebaier, and M. Takizawa : A Model for


Reducing Power Consumption in Peer-to-Peer Systems, IEEE
Systems Journal, Vol.4, Issue.2, May 2010, page:221-229.

[4] T. Enokido, K. Suzuki, A. Aikebaier, and M. Takizawa


: Process Allocation Algorithm for Improving the Energy
Efficiency in Distributed Systems, Proc. of IEEE the 24th
International Conference on Advanced Information Network-
ing and Applications (AINA2010), Perth, Australia, 2010,
page:142-149.

[5] T. Enokido, A. Aikebaier, and M. Takizawa : Process Allo-


cation Algorithms for Saving Power Consumption in Peer-to-
Peer Systems, accepted for publication in IEEE Transactions
on Industrial Electronics, 2011

[6] T. Enokido, A. Aikebaier, and M. Takizawa : Process Allo-


cation Algorithms for Saving Power Consumption in Peer-to-
Peer Systems, accepted for publication in IEEE Transactions
on Industrial Electronics, 2011.

[7] “Watts up?”, https://www.wattsupmeters.com/secure/


products.php?pn=0wai=0soft=4, (10-11-2010)

[8] A. Ghobadi, C. Eswaran, N. Muthuvelu, I. K. T. Tan, and Y.


L. Kee : An adaptive wrapper algorithm for file transfer appli-
cations to support optimal large file transfers, Proc.IEEE the
11th International Conference on Advanced Communication
Technology (ICACT 2009) , 2009

617

View publication stats

You might also like