Accepted Manuscript

Title: Experimental investigations of air conditioning solutions in high power

density data centers using a scaled physical model

Author: S.A. Nada, K.E. Elfeky, Ali M.A. Attia

PII: S0140-7007(15)00322-9
DOI: http://dx.doi.org/doi: 10.1016/j.ijrefrig.2015.10.027
Reference: JIJR 3183

To appear in: International Journal of Refrigeration

Received date: 12-8-2015

Revised date: 21-10-2015
Accepted date: 23-10-2015

Please cite this article as: S.A. Nada, K.E. Elfeky, Ali M.A. Attia, Experimental investigations of
air conditioning solutions in high power density data centers using a scaled physical model,
International Journal of Refrigeration (2015), http://dx.doi.org/doi:
10.1016/j.ijrefrig.2015.10.027.

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will
undergo copyediting, typesetting, and review of the resulting proof before it is published in its
final form. Please note that during the production process errors may be discovered which could
affect the content, and all legal disclaimers that apply to the journal pertain.
Experimental investigations of air conditioning solutions in high power density data centers
using a scaled physical model
S.A. Nada*, K. E. Elfeky, Ali M.A. Attia

Department of Mechanical Engineering, Benha Faculty of Engineering, Benha University, Benha, Egypt.
*Corresponding author (Sameh Nada). Email: [email protected], Mob: +201066611381

Highlights
Different solutions of thermal management of data centers are investigated, compared and
evaluated.
A scaled physical model data center have been designed and constructed using the theory of
scale modeling.
Effects of using aisle partition and aisle containment on rack inlet temperature are studied.

Abstract

The widespread use of data centers, the dramatically increase of the data center power density

and the need of improving cooling system efficiency to maintain reliable operation temperature

and save cooling energy make the study of data center thermal management an urgent issue. In

the current paper, three different configurations for thermal management solution of high power

density data centers are investigated, compared and evaluated. A scaled physical model data

center have been designed and constructed for the sake of the study using the theory of scale

modeling of air flow experiments. The results showed that (i) by using aisle partition and aisle

containment the rack inlet temperature can be reduced by 3–13% and 13%-15.5% for aisle

partition and aisle containment configurations, respectively, (ii) the intake air temperature

reduction increases with increasing power density, (iii) using aisle partitions and aisle

containment with raised floor improves the data center cooling performance.

Page 1 of 27
Keywords

Data center, air conditioning, physical modeling, thermal management, aisle partition, aisle

containment.

Nomenclature
Q heat dissipation (W)
Cp specific heat of air at constant pressure, (J kg-1 k-1 )
mass flow rate (kg s-1 )
T temperature (◦C)
Tref reference temperature (◦C)
CRAC Computer Room Air Conditioning
RHI return heat index
SHI supply heat index
U velocity (m s-1)
L length (m)
ν kinematic viscosity (m2 s-1)
α length scale = Lm/LR
τ time scale
Re Reynolds number
Ar Archimedes number
Pr Prandtl number
Superscripts
r Rack
c CRAC
Subscripts
in Inlet
out Outlet
m for model
R for real data center

Page 2 of 27
i,j Cartesian direction

1. Introduction

Data centers are widely used in different industrial applications where large/high-speed data

processing is necessary, such as telecommunications, data storage and processing in banks,

market transactions and others special and private applications. Recent studies showed that data

center consumes a huge amount of the total power consumption of modern cities. It was reported

that data centers consumed 61 billion kWh or about 1.5 % of U.S. total electricity consumption

in 2006 [1]. A large portion of this consumed energy (almost 50%) is necessary for cooling of

servers to maintain their temperature within the allowable limits [2]. The properly manage data

center cooling process would reduce this portion of energy. Consequently, a much more detailed

understanding of air flow and temperature distributions for proper thermal management in data

centers is a vital issue to operate the data centers within the required specifications while

avoiding excessive use of cooling. Layout and features of all data centers are similar; most they

use raised-floor configuration. Figure 1 shows a typical schematic view of open aisle data centers

[3]. The racks are arranged in a hot-/cold-aisle configuration with standard alignment like that

shown in Table 1 [4]. The cold aisle contains perforated tiles that supply cold air to the inlets of

the server racks from the under-floor plenum. The hot air leaving the racks is extracted by the

Computer Room Air Conditioning (CRAC) unit to re-cool and supplies it as cold air to data

center plenum to complete the cycle. This concept of energy management for data centers

prevents the unwanted mixing of the hot air expelled from the servers with the cooling air

coming from the perforated tiles. However, hot air recirculation and cold air bypass must be

considered in design and operation stage in order to prevent a drop in the servers cooling

efficiency.

Page 3 of 27
Efficient thermal management of data centers can be maintained by using proper air distribution

in the room that would reduce or prevent the hot air recirculation and/or the cold air bypass. For

this purpose, physical separation (aisle partition) of hot and cold aisles has been suggested [5].

Containment of air throughout the data center is an important thermal management and energy

saving strategies that results in the data center optimization especially in high power density data

center. Most of the modern energy efficient data centers use some kind of containment system

[6-9]. Generally, the major benefit of aisle partition and containment is the mitigation of server

air inlet temperatures due to the minimum mixing of cold air with hot air. Data center thermal

management performance and effectiveness is normally evaluated in terms of performance

dimensionless metrics; Supply Heat Index (SHI) and the Return Heat Index (RHI) [10, 11].

Using these indices, heat transfer and thermal management inside the data centers can be

understood and evaluated.

One of the relevance work in this area is the work of Cho et al. [12] who studied air distribution

inside high compute density (Internet) data centers. Cho et al. [12] observed that the air velocity

is not an important factor for the data center designers where human thermal comfort is not a

significant factor. Shrivastava et al. [13] reported that supply the cold air from raised floor and

extract the return air from ceiling is the most efficient air distribution system. On the other hand,

the ceiling supply with under floor return leads to the worst air distribution and the worst thermal

managements in data centers [14]. Other works cited in the literature [15-16] to evaluate and

compare the under floor supply and the overhead supply configurations. They reported that

although under floor supply is recommended for proper air distribution and thermal

Page 4 of 27
managements it can result in hot spots at the servers located at the rack top due to hot air

recirculation.

More recently, Van Gilder et al. [17] studied the uniformity of air flow through the raised floor

perforated tile and they reported that perforated tiles of 25% opening ratio gives the best flow

uniformity. Karlsson and Moshfegh [18] experimentally studied the temperature distribution at

racks inlet using infrared cameras. They reported that a temperature gradient exists along racks

height, where the rack top has higher inlet temperatures. Patterson [19] studied the effect of data

center temperature on energy efficiency. It was reported that the efficiency of the overall cooling

system is strongly related to the efficiency of the data center room cooling which in turn depends

on the data center air temperature. Durand-Estebe et al. [20] studied the simulation of a

temperature adaptive control strategy in a data center. It was reported that the increase in blown

air temperature allows the system to raise the evaporator loop temperature reducing the load on

the heat pump. It was found that the servers reduces its energy needs by about 78% when the

blow air temperature ( CRAC output air temperature ) rises from 16 °C to 24 °C. Nada et al. [21]

Studied the effect of the schemes of server’s power loadings on data center thermal

management. It was found that uniform power loading provides the best results for (SHI/RHI)

and the optimum benefit of cold air in server and clustering of active servers lead to better

air flow management compared to discretely individual active servers and segmented

distributions of active servers.

Page 5 of 27
Cho et al. [5] observed that, the data center cooling efficiency can be improved by installing a

simple partition wall on the rack server. Gondipalli et al. [6] conducted a computational study on

the effect of using cold aisle compartments. The results showed that a 15-40% reduction in rack

inlet temperature can be obtained by using cold aisle compartments for the same room layout and

cold air supply. In this regard, Vikneshan et al. [7] reported that, the fully provisioned contained

aisle is preferred in case of no geometrical or cost limitations otherwise the partial containment is

a good accepted option. Vaibhav K. et al. [8] showed that, the containment of the cold aisle tends

to significantly improve the temperature uniformity in the cold aisle, as well as at the server

inlets temperature. Saurabh K. et al. [9] found that, a containment system reduces the overall

cooling energy cost by preventing or reducing the mixing of cold and hot air streams.

Actually, most of these investigations were conducted on a real data center which is not an easy

task due to high cost and difficulty to be controlled. Fernando et al. [22], studied the viability of

design and construct a scaled model for the purpose of testing an actual data center using the

theory of scale modeling for airflow experiments. Results showed accurate thermal similarity

while the airflow similarity cannot be obtained with reasonable accuracy.

In the present work, we focus on the thermal characteristics of separation and contained cold

aisle by make comparison between three air distribution system configurations; typical under

floor air cooling system configuration, typical configuration with aisle partition system and

typical configuration with aisle containment. It is important to investigate the influence of these

air distribution system configurations on the rack temperature distribution, the server temperature

distribution, and the performance metrics (SHI and RHI) at different power densities. For

purpose of study, a physically modeled data center with a single rack and four server simulators

Page 6 of 27
is constructed following the under floor supply-ceiling return airflow entering through perforated

plate of 25% opening ratio.

2. Experimental facility and procedure

2.1. Scale modeling of air flow and temperature.

The above literature revealed that a scale modeling theory is needed to achieve a physical scale

and air flow modeling of real data centers for the purpose of testing and study [23]. The proposed

model is based on the use of under floor supply-ceiling return configuration as it provides the

best cooling performance [13-16]. From the limited available literature regarding the data center

modeling, there are three papers; one for testing the scaled model of data centers [22], the second

for testing the air flow in the scale modeled room [24] and the third for testing the effect of the

rack intake temperature on the energy consumptions of the computer room air conditioning unit

[20]. Flow conditions are completely similar if all relevant dimensionless parameters have the

same corresponding values for the model and the real data centers. It was verified that [22, 24],

such similarity can only be achieved if geometric, kinematic and thermal similarities between the

modeled and real data centers exist. From dimensional analysis, this complete similarity can only

be attained if Prandtl (Pr – ratio of momentum diffusivity to thermal diffusivity), Reynolds (Re –

ratio of momentum force to viscous force), and Archimedes (Ar - ratio of gravitational force to

viscous force characterizing fluid motion due to density differences) numbers equal for both

modeled and real data centers while geometric and boundary conditions similarities are attained.

In fact achieving complete similarity is a very difficult, so it is often more reasonable to try to

achieve particular type of similarity: geometric, kinematic, dynamic, and/or thermal similarities

[23]. Assuming that geometric and boundary similarities exist then equality of Pr, Re and/or Ar

Page 7 of 27
must be obtained to achieve particular types of similarities. Applying this similarity criterion to

the supply air from perforated tile means:

1- equality of Prandtl number is achieved when the same working fluid (i.e. air) in the

scaled model and real data centers is used.

2- equality of Reynolds number (Re) for both model and real data centers while using the

same working fluid (i.e. identical kinematic viscosity or ) provides:

then a time scale factor can be defined as:

3- equality of Archimedes Number (Ar) for model and real data centers is given by:

Where g is the gravitational acceleration (9.81 m/s2), β is the expansion coefficient

( for air at 20°C), and ΔT is the temperature rise throughout room

(typically, air going through a server increases 10°C in temperature), L is a characteristic

vertical length scale (2 m standard rack height), and U is a characteristic velocity

(average tile flow velocity). It is common in real data centers to have value of Ar greater

than 1, thus natural convection dominates the data center cooling and so buoyancy must

be taken into account.

Page 8 of 27
 In scaling velocity, a fully open perforated tile velocity is considered. If the temperature

change is similar between the model and real data centers, then the time-scale ratio based

on Archimedes similarity gives:

It can be seen that the time-scale ratios based on Re equality (in Eqn. 2) and that based on Ar

equality (in Eqn. 4) do not agree unless the length-scale ratio is 1. Therefore, a completely

similar model of a data center is impossible to be created. For non-isothermal flows, Awbi and

Nemri [24] recommend that the time-scale factor based on the equality of Ar and Re can be used

when Archimedes number is greater than and less than 0.001, respectively

The experimental facility of the present work is constructed using a scaled model of a full size

standard dimension data center based on a length-scale ratio of α equals 1/6. Using this length

scale ratio and take typical values of real rack height and tiles air velocity as LR = 2 m and UR =

1 m/s, respectively, then Re and Ar for the scaled model data center can be estimated as.

Where Um is the average inlet velocity of the air jet from the tiles (calculated from Eqn. 2 as Um

= UR /α = 6 m/s for Re equality and from Eqn. 4 as Um = UR =0.4082 m/s for Ar equality)

and Lm is a characteristic vertical length scale (Lm = α LR = 0.334 m), i.e.

Page 9 of 27
As is greater than 0.001, then the time-scale factor based on Ar equality will be used

to achieve the similarity analysis in the present study.

Considering the one-sixth geometrical scaled model and a time scale factor based Archimedes

number equality, Table 2 summarizes the modified boundary conditions, flow rates and heat

generation for the scaled model data center. The table also gives the necessary equations used to

study heat transfer in the modeled data center. The heat generation is determined from the server

flow rate and prescribes the temperature difference of the air flowing through the server (
◦
C [25]).

2.2. Experimental setup

A physical scale room model was used to conduct the experimental investigation of the present

work to avoid the required exhaustive construction, measurements costs and time-length efforts

of actual data centers. Scale modeling theory was utilized to design and construct the scale room

model as per item 2.1. A test facility including a scaled data center room, a rack of servers inside

the scaled room and cooling air supplying circuit was designed and constructed to simulate the

conditions and arrangements of actual standard data centers. The built-up test facility was

equipped with different measuring devices and instrumentations to measure different parameters

needed for later analysis such as temperatures, powers, and air flow rates. Figure 2 shows a

schematic diagram of the whole test facility and measuring instruments. As a scenario of

operation, the blower delivers air into a space that simulates the raised floor of actual data center.

This air enters the data center room through perforated tiles and passes through the front face of

the data center rack to cool its servers. The hot air is sucked to flow throughout the cooling

passage of server with the help of specific fan. The hot air exits from the rear face of the rack and

is discharged to the atmosphere at the top of the scaled room using discharging fan. In the current

10

Page 10 of 27
test facility, a single rack including four servers is sued to simulate the actual rack of a real data

center located at the center of the room and surrounded by hot and cold aisles. To measure

temperature distribution at different locations within scaled data center, number of

thermocouples (type T) were installed; typically twenty eight thermocouples. Plastic frames were

used to fix the thermocouples on rack inlet and exit to measure the corresponding air temperature

distribution. The thermocouples were mounted at 2 cm in the front and back of the air intake and

exit rack door. Each frame contains eight thermocouples distributed on it at different heights.

Two sets of two thermocouples are installed underneath the perforated tile and on the return fan

intake to measure the supply and exit air temperatures. The analog signal of all thermocouples

has been converted into digital values and saved in Excel data sheet for later analysis via Data

Acquisition (DAQ, model NI cDAQ-9178 with NI 9213 16-Channel Thermocouple Input

Module) connected with PC and controlled by LabView program. Reading of all thermocouples

received by DAQ has been corrected against calibrated thermometer from normal temperature up

to 150 °C.

2.3. Scale Model Overall Design

The data center room used in the current experimental facility is a scale model of a full size

standard dimension data center room with a length-scale ratio α=1/6 based on Archimedes

number equivalence, see Figure 3 for more details. It is made from Plexiglas wall of thickness

1cm, air tight assembled using silicon. The room dimensions is 400 x 329.5 x 500 mm. The

raised floor thickness of the room is 100 mm. The cold ails and hot ails dimensions are 101.6 and

75 mm respectively.

11

Page 11 of 27
A single rake of dimensions 101.6 x 152.6 x 334 mm (height) located at the center of the data

center room was designed and built to accurately simulate a rack in actual data center located a

the middle of racks matrix. In order to simplify the modeling, the rack was designed to house

four servers accommodating four server simulators [26]. To ensure that there is no internal

recirculation, server intake and exhaust face is attached to the rack perforated doors made of

screen mesh of 65% opening ratio simulating actual servers opening ratio. The dimensions of

each server cabinet simulator is 101.6 mm wide and 80 mm high and 152.6 mm deep. Each

server has a variable speed fan (up to ~ (0.45 /min) and electric heater of variable heating

power (up to ~ 150 W) simulating the fan and heat generation of actual servers. The advantage of

using these server simulators rather than actual servers is the ability to quantify the controlling

parameters such as fan speed and heat dissipation .The flow rate for the server fan is controlled

by changing the supplied power using Variable AC transformer (variac - that regulate voltage in

the range of 0 – 220 volts). The fans flow rates are measured by using hot wire anemometer

(Model: Testo 435, of measuring range 0 to 20 m/s). Heat is generated in each server by using a

nickel-chromium wire wrapped on a plate of mica (electrical insulation and not thermally

insulated) covered by layer of stainless steel. Figure 4 shows a top view of the server heater. In

order to obtain a uniform surface temperature a 0.5 mm in thick stainless steel plate was attached

to the outer surface of the heater. The input power to server was controlled using variac. Two

thermocouples located on the heater surface of each server (see Fig. 4) are used to measure the

temperature of the heaters.

12

Page 12 of 27
2.4. Experimental conditions.

The experiments were conducted at different operating and geometric parameters to study the

effects of these parameters on the thermal and air flow management inside the scaled model data

center. The studied parameters and ranges include:

Room power density (W/m2): 379, 759, 1139, 1518 and 1898

Blower air discharge temperature: 22 °C

Data center system configuration: Typical under floor, Typical under floor with aisle partition

and Typical under floor with aisle containment system.

The room power density is the sum of active servers’ power per unit area of the data center

room. The comparison involves five heat densities (379 W/m2, 759 W/m2, 1139 watt/m2, 1518

W/m2 and 1898 W/m2) and three room configuration (open, aisle partition, and aisle

containment). The open configuration does not use physical barriers to enhance the separation of

hot and cold air in the equipment room (see Fig, 5(a)). The aisle partition system vertically

dividing the cold aisle and the hot aisle with height 100 mm , as shown in Fig. 5(b), and the aisle

enclosure system that blocks off the upper part of the cold aisle, as shown in Fig. 5(c). Fig.5

shows side views of the three room configuration. All experiments were conducted at 25%

perforated plate opening ratio as this ratio is the most widely used in data center applications and

is also recommended by previous investigators [17, 26-28] who reported that 25% opening ratio

gave the more uniform air flow distribution and bitter supply heat index along the rack servers as

compared to other opening ratios. This opening It is important to state that the power supplied to

blower is synchronized with the predefined server power, as actually done in real data centers to

13

Page 13 of 27
maintain server temperature within specific range of operating temperature.

2.5. Experimental procedure and program.

The procedure and experimental program were performed as follows:

1. Make sure the room is clean and accessible before starting the experiment.

2. Make sure that blower is operational and adjust the blower fan speed according to the

experimental program.

3. Supply and adjust power to each server in the rack according to the program.

4. Turn on the server fans and adjust the fan speed according to the experiment objective.

5. Turn on the data acquisition system.

6. Wait until steady state condition is achieved. Steady state conditions was assured when

the heater surface temperature of each server was maintained constant (variation within

±0.2 °C for 5 min). Steady stat conditions was normally achieved after about forty

minutes.

7. Measure the tile flow rates as well as all temperature values.

8. Record the readings of all instruments (voltage, current, flow rate and temperatures).

Repeat steps 3–8 at different power density for specific data center configuration systems. Then these

steps are repeated at other configurations of data center system in accordance with experimental

program. Each experiment is repeated twice to ensure the consistency in measurements. The

quantities measured directly in each experiment include air flow rate, air temperatures, input

voltage and input current. The uncertainties in measuring these quantities were evaluated to

be ±2%, ±0.2 °C, ±0.25% and ±0.25%, respectively.

14

Page 14 of 27
2.6. Data reduction and thermal metrics for data centers

Thermal metrics are used to evaluate the data center airflow performance and thermal

management. In a real data centers recirculation, bypass or infiltration phenomena may occur

(see Figure 6). Accordingly, airflow ingested by a rack is a mixture of cold and hot airflow

coming from surrounding. In case of recirculation, the exhaust hot air travels back into the

server’s intake air stream leading to rise of intake air temperature and so to unsafe servers

operating temperatures. In case of bypass airflow, the supply cold source air bypasses the active

servers and moves directly into the hot exhaust air stream. It has been reported that only about

40% of supplied air passes through the servers due to the occurrence of bypass and/or leakages

[24]. In general, thermal metrics depend on the geometric and physical parameters of the data

centers.

To evaluate cooling performance of data centers, Sharma et al. [6] proposed dimensionless

parameters for thermal indices. These indices evaluate the extent of cold and hot air mixing in

data center. They are scalable metrics and potentially applicable at racks, rows or at wide data

center level. Supply heat index parameter is defined as the ratio of the heat gained by cold aisle

air before it enters the racks to the heat gained by the air leaving the racks and Return heat index

parameter is defined as the ratio of the total heat extraction by the CRAC units to the heat gained

by the air leaving the racks. The utilization of dimensionless parameters allows these formulas to

be scalable for any size system. These indices are defined as:

15

Page 15 of 27
Where Q is the total heat dissipation from all the racks in the data center, Q is the enthalpy rise

of the cold air before entering the racks, is the mass flow of air through the rack in the

row of racks, and are the average inlet and outlet temperature from the

rack in the row of racks and is the vent tile inlet air temperature (assumed to be

identical for all rows). Neglect heat transfer in the plenum, the temperature of the air exit from

the vent tile and CRAC supply air temperature are considered to be equal and denoted by the

reference temperatures in enthalpy calculations. According to mass conservation at rack inlet and

exit, SHI can be rewritten as a function of rack inlet, rack outlet and CRAC outlet temperatures.

For a single rack data center, these metrics are evaluated using the following relations:

Equations 7 and 9 reveals that higher leads to higher and hence a higher SHI. When

the inlet temperature ( ) to the rack rises, systems failure is expected and reliability problems

exist. Increasing increases entropy generation due to mixing and this reduces energy

16

Page 16 of 27
efficiency for the data center. Therefore, SHI can be an indicator of thermal management and

energy efficiency in data center.

A low RHI indicates the mixing of rack exhaust air with the cold air due to bypass air flow.

Target values of SHI and RHI are 0 and 1; respectively and typical benchmark of recommended

acceptable ranges of SHI and RHI are SHI<0.2 and RHI>0.8.

3. Results

To study the effect of using the three air distribution system configuration on thermal

management (temperature distribution, and SHI and RHI ) at different power densities, the

experiments were conducted at different power levels in the room from 379 to 1898

by step 380 for a typical under-floor air cooling system of 25% perforated tile opening

ratio and uniform servers power schemes configuration. The four servers are equally powered

and the speed setting of their fans are identically set to discharge the sum of uniform air flow rate

of 0.0042 m3/s across the entire rack for the first case. The results shown in figs. 7-11 provides

the rack temperature profile and servers’ surface temperature distribution at the studied range of

power density for the three configurations superimposed on the same plots for the sake of

comparison. For all the three configurations, temperatures are monitored along the rack at

different heights above the raised floor and at the surfaces of the servers. Average of these

temperatures is used as a comparison parameter for aisle partition and aisle containment with

typical under floor air cooling system configuration.

As observed from Figs. 7-11, for the three configurations there is a significant increase of the

temperature at the rack back and the servers’ surface temperatures with the increase of the power

density. This can be attributed to fact that, increasing the power density increases the air flow

17

Page 17 of 27
rate supplied from the tile and this increases the cold air velocity discharged from the tiles which

leads to lower static pressure at the cold ails. Decreasing the cold ails static pressure decreases

the air flow rates of the server fans and this leads to the increase of the server’s surface

temperature and the exit fans temperature (temperature at rack back). It is also noticed that, the

server located at the rack bottom cabinet always has lower temperature compared to the upper

servers no matter the value of the power density. Moreover, along the height of the rack, there is

a remarkable increase in the server’s surface temperature reaching the maximum value at the

highest location; server 4 (located at height 25 cm). The increase of the surface temperature of

servers with the increase of its location height can be attributed to the buoyancy effect that makes

the environment of the server at higher levels in the rack hotter than those at lower levels of the

rack.

The discussion of the behavior in the previous section reveals that increasing the perforated tile

flow rate (configuration 1) by using active tiles (tiles with fans to increase the air flow rate) will

not resolve the increase in the needed cooling demand. In fact, increasing the perforated tile

supply flow rate would adversely impact rack cooling due to the increase of the server’s

temperature. Techniques to prevent re-circulation and by-pass in case of high power density is

needed. Two alternative techniques are used; aisle partition system configuration and aisle

containment system configuration to contain the air in cold aisles and hot aisles and prevent the

by-pass or the recirculation.

From Figures 7-11(a), the average inlet temperature is improved no matter the value of the power

density due to the existence of aisle partition as aisle partition prevents air re-circulation to some

extent in comparison with the case where no partition is used. In this case, the inlet average

temperature drops from 26°C to 22.6 °C showing an improvement of 13% at 1898 power

18

Page 18 of 27
density for aisle partition. The results presented in Figs. 7-11 for aisle partition are in a good

agreement with that obtained by Cho J. and Kim B [5]. A similar effect has been found for the

average surface temperature of servers (see Figure 7-11(b)), for example the top server’s surface

temperature is improved by 11% at 1898 power density. This effect of aisle partition

regarding the surface temperature is not significant for bottom servers where the cooling air

temperature is already low and the uncontrolled phenomenon have low tendency to occur. But at

high altitude, the buoyance effect of natural convection makes hotter gases at the top leading not

only to higher surface temperature of servers at the top, but also increase the tendency for mixing

both hot and cold air above the top server.

By installing cold aisle enclosure (cold aisle containment) instead of aisle partitions, the average

inlet temperature is additional improved for all studied range of power densities (see Figure 7-

11(a)) because the aisle enclosure completely block the top recirculation. The inlet average

temperature drops from 26°C to 22.1°C for 1898 power density for aisle enclosure .In this

case, greater enhancement of servers’ cooling is attained for all servers along the height of the

rack with overall improvement around 15.5%. The results presented in Figs. 7-11 for aisle

enclosure are in a good agreement with that obtained by Gondipalli et al [6]. Generally, the

improving effect in thermal management when aisle enclosure is used primarily comes from the

full obstruction of the top circulation.

19

Page 19 of 27
Figure 12 shows the air-management and energy performance as expressed by SHI and RHI at

different power densities for the studied three configurations. Generally as the power density is

increased, there is a remarkable decrease in the values of SHI and a remarkable increase in the

values of RHI, which directly reflects the improving of heat transfer and cooling efficiency for all

studied air distribution system configurations. A slight improve in the thermal management

efficiency is received when aisle partition is used and this effect is increased with the increase of

power density (SHI is improved from 13% to 62%, and RHI is improved from 4% to 7% as power

density varied from 379 to 1898 W/m2). On the other hand the best improvement in cooling

efficiency is attained when aisle enclose is used. This enhancement effect due to the use of aisle

enclosure is increased (over the case where no enclosure or partition were used) with the increase

in power density (SHI is improved by a fixed value of order 70%, and RHI is improved from 8%

to 25% as density varied from 379 to 1898 W/m2). Even at the 1898 watt/m2 power level, the RHI

is elevated approximately to 1 (ideal). Many data centers that are running out of capacity due to

limited raised-floor heights. Raised-floor heights can be reconfigured with cold-aisle containment

to allow very significant heat densities. This should provide a welcome relief for many data

centers. Finally, it should be noted that the aisle containment may be more vulnerable to

catastrophic cooling outages, and engineered thermal solutions are often required to safeguard the

servers.

Conclusion

20

Page 20 of 27
In this paper effect of three air distribution system on rack temperatures profile and thermal

management have been studied in detail by using two performance metrics for analyzing air-

management systems in data centers. Two performance :The Supply heat Index (SHI) and the

Return heat Index were computed experimental to evaluate the thermal performance of the scaled

model data center at five different power densities and three room architectures. Bases on the

dimensional analyses, a scaled data center of one rack accommodating four servers was designed

and constructed based on a scale ratio of 1/6. Front and rear rack temperatures distributions,

server’s temperatures and supply and return heat indices are measured and used to study,

compare and evaluate the performance of the three air distribution system configuration in the

scaled data center and its efficiency in simulating actual data centers. Efficient thermal

management of data centers can be improved by using proper isolation technique in data center

.Both isolation techniques aisle partition and aisle containment reduce the rack inlet temperatures

at critical locations (top portions of server racks) significantly. Improvement in thermal

management primarily comes by obstructing top circulation. Isolation technique with just aisle

partition has marginal merits and may improve the thermal performance of top server racks as

aisle partition prevents air re-circulation to some extent in comparison with the existing system.

But blocking the cold aisle give the best thermal performance because the aisle enclosure block

top recirculation completely.

The results showed that.

 Both isolation techniques aisle partition and aisle containment reduce the rack inlet

temperatures at critical locations (top portions of server racks) significantly. Isolation

technique with aisle partition and aisle containment can reduce the rack inlet temperature

21

Page 21 of 27
 by ~ 3%– 13% for aisle partition and by ~ 13%– 15.5% for aisle containment at different

power density.

 Isolation technique with just aisle partition improve the thermal performance of top

server racks only.

 Blocking the cold aisle give the best thermal performance because the aisle containment

block top recirculation completely.

Reference.

[1] EPA. Report to Congress on Server and Data Center Energy Efficiency, Public Law 109-

431, 2007, USA.

[2] ASHRAE. Environmental Guidelines for Datacom Equipment, ASHRAE, Atlanta, GA, 2008,

USA.

[3] Patankar S. V. Airflow and cooling in a data center. Journal of Heat Transfer, 2010; 132 (7):

073001.

[4] TC 9.9, Mission Critical Facilities, Technology Spaces, and Electronic Equipment, Thermal

guidelines for data processing environments. American Society of Heating, Refrigerating and

Air-Conditioning Engineers Inc., 2004.

[5] Cho J., Kim B. Evaluation of air management system’s thermal performance for superior

cooling efficiency in high-density data centers. Energy and Buildings 2011; 43: 2145–2155.

[6] Gondipalli S., Bhopte S., Sammakia B., Iyengar M., Schmidt R.R. Effect of Isolating Cold

Aisles on Rack Inlet Temperature. Proceeding of the 11th IEEE ITHERM Conference 2008,

Orlando, FL, USA.

22

Page 22 of 27
[7] Vikneshan S., Vaibhav K. A., Yogendra J. Experimental characterization of cold aisle

containment for data Centers. 29th IEEE SEMI-THERM Symposium, Atlanta, GA 30332, USA,

2013.

[8] Vaibhav K. A., Vikneshan S., Yogendra J. thermal characteristics of open and contained data

center cold aisle. ASME Journal of Heat Transfer UNE2013, Vol. 135 / 061901-1.

[9] Saurabh. K. S., Andrew. R. C., Mahmoud. I. quantitative comparison of air containment

systems. Proceedings of the Intersociety Conference on Thermal and Thermomechanical

Phenomena in Electronic Systems (ITherm). San Diego, CA, 2012, May 30-June 1.

[10] Herrlin M.K. A new tool for evaluating and designing the thermal environment in telecom

central offices. Proceedings of Telecommunications Energy Conference 2006; 1–5.

[11] Sharma R.K., Bash C.E., Patel C.D. Dimensionless parameters for evaluation of thermal

design and performance of large scale data centers. Proceedings of AIAA2002-3091, American

Institute of Aeronautics and Astronautics Conference, 2002.

[12] Cho J., Lim T., Kim B. S. Measurements and Predictions of the Air Distribution Systems in

High Compute Density (Internet) Data Centers. Energy and Buildings 2009; 41(10): 1107-1115.

[13] Shrivastava S., Sammakia B., Schmidt R., Iyengar M. Comparative analysis of different

data center airflow management configurations. Proceedings of the ASME/Pacific Rim

Technical Conference and Exhibition on Integration and Packaging of MEMS, NEMS, and

Electronic Systems: Advances in Electronic Packaging 2005, PART A: 329-336.

23

Page 23 of 27
[14] Nakao M., H. Hayama, and M. Nishioka. Which cooling air supply system is better for a

high heat density room: Underfloor or overhead. Proceedings of International

Telecommunications Energy Conference (INTELEC) 1991; 12 (4): 393- 400.

[15] Sorell, V., S. Escalante, and J. Yang. Comparison of overhead and underfloor air delivery

systems in a data center environment using CFD modeling. ASHRAETransactions 2005;

111(2):756-64.

[16] Herrlin, M., and Belady, C. Gravity assisted air mixing in data centers and how it affects the

rack cooling effectiveness. Proceedings of the Inter Society Conference on Thermal Phenomena

(ITherm), San Diego, CA, 2006 May 30-June 2.

[17] VanGilder J., Schmidt R. Airflow uniformity through perforated tiles in a raised floor data

center. Proceedings of the ASME InterPACK '05 Conference, San Francisco, CA, paper IPACK

2005-73375.

[18] Karlsson J.F., Moshfegh B. Investigation of indoor climate and power usage in a data

center .Energy and Buildings 2005; 37: 1075–1083.

[19] Patterson MK. The Effect of data center temperature on energy efficiency. In: Intel

Corporation IEEE; 2008, 978-1-4244-1701-8.

[20] Estebe, B.D., Bot, C.L, Mancos, J.N., Arquis, E., study the Simulation of a temperature

adaptive control strategy for an IWSE economizer in a data center", Applied Energy, Volume

134, 1 December 2014, Pages 45-56.

[21] S.A. Nada, K. E. Elfeky, Ali M.A. Attia, W.G. Alshaer, Thermal Management of Electronic

Servers under Different Power Conditions, International Journal of Emerging Trends in

Electrical and Electronics (IJETEE) Vol. 11, Issue. 4, Aug 2015.

24

Page 24 of 27
[22] Fernando H., Siriwardana J., Halgamuge S. Can a Data Center Heat-Flow Model be Scaled

Down?" Information and Automation for Sustainability (ICIAfS) IEEE, 2012: 273 - 278.

[23] F. M. White, Fluid Mechanics, 4th ed. McGraw Hill, New York, 2001.

[24] Awbi H., Nemri M. Scale effect in room airflow studies. Energy and Buildings 1990; 14(3):

207 – 210.

[25] Rasmussen N. Cooling strategies for ultra-high density racks and blade servers, American

Power Conversion, Washington, DC, White Paper.

[26] Smith J. F., Abdelmaksoud W.A., Erden H.S., Dannenhoffer J. F., Dang T.Q., Khalifa H.E.,

Schmidt R. R., and Iyengar M. Design of Simulated Server Racks for Data Center Research.

Proceedings of ASME InterPACK, Portland, OR, 2011 July 6 – 8.

[27] Radmehr, A., Karki,K.C., and Patankar,S.V. Analysis of airflow distribution across a front-

to-rear server rack. Proceedings of the ASME InterPACK '07 Conference, Vancouver, British

Columbia, Canada, paper IPACK2007-33574.

[28] Abdelmaksoud, W., Khalifa, H., E., Dang, T., Iyengar, M., and Schmidt, R., 2010,

“Experimental and Computaional Study of Perforated Floor Tile in Data Centers,” Proceedings

of the Inter Society Conference on Thermal Phenomena (ITherm), Las Vegas, Nevada, USA.

25

Page 25 of 27
 Figure 1: typical open aisle data center [3]

Fig. 2. Schematic diagram of the experimental setup

Figure 3. Scale model data center (a) top view, (b) side view (dimensions in millimeters)

Fig.4. Shows a top view of the server (dimensions in millimeters).

Fig. 5. Comparison of three air distribution system. (a) Typical under floor air cooling system
configuration, (b) typical configuration with aisle partition system, and (c) typical configuration
with aisle containment system.
Fig. 6 Typical data center airflows. (a) Re-circulation airflow and (b) by-pass airflow.

Fig.7. Comparison of the different configurations systems at 379 watt/m2 power density,(a)

Temperature profile at front and rear of rack for and (b) Servers’ surface temperature

distribution.

Fig.8. Comparison of the different configurations systems at 759 watt/m2 power density,(a)

Temperature profile at front and rear of rack for and (b) Servers temperature distribution.

Fig.9. Comparison of the different configurations systems at 1139 watt/m2 power density, (a)

Temperature profile at front and rear of rack for and (b) Servers temperature distribution.

Fig.10. Comparison of the different configurations systems at 1518 watt/m2 power density,(a)

Temperature profile at front and rear of rack for and (b) Servers temperature distribution.

Fig.11. Comparison of the different configurations systems at 1898 watt/m2 power density, (a)

Temperature profile at front and rear of rack for and (b) Servers temperature distribution.

Fig. 12. Variations of SHI and RHI with power density for the three air distribution system

configuration.

26

Page 26 of 27
Table 1: Aisle pitch allocation and rack arrangements with separation [4].

Table 2: Boundary conditions and necessary calculations of sever load of modeled data center

Boundary condition Symbol Equation Value

Inlet airflow velocity Eqn. 4 0.4083 ms−1
Tile flow rate = 4.215* m3s−1
Inlet air temperature - 22 °C
Rack Temperature rise - Assuming (∆T=10 °C)
Server flow rate = /4 1.0534* m3s−1
Server heat generation 12.969 W
Tile porosity - 0.25
Rack porosity - 0.65

27

Page 27 of 27