Content uploaded by Monem Beitelmal

Author content

All content in this area was uploaded by Monem Beitelmal on Jan 26, 2017

Content may be subject to copyright.

ASME 2015 International Mechanical Engineering Congress & Exposition

November 13-19, 2015, Houston, Texas, USA

IMECE2015-50884

NUMERICAL INVESTIGATION OF DATA CENTER RAISED-FLOOR PLENUM

Abdlmonem H. Beitelmal

Qatar Environment and Energy Research Institute

Qatar Foundation

Doha, Qatar

ABSTRACT

Data center raised-floor plenum effectiveness is

numerically investigated using a computational fluid

dynamics (CFD) package to determine the most

appropriate data center raised floor plenum height (size).

The current study considers raised floor plenum height

between 30.5 cm and 152.4 cm (12-60 inch) with the

standard 15.2 cm (6-inch) increment while maintaining

the supply airflow rate constant. Three factors are

considered for optimum plenum size: the individual

airflow rate from each perforated tile, the level of

airflow rates uniformity between different perforated

tiles and the top rack inlet air temperature at 2.13 m (7-

ft) above the raised floor. The results show that raised

floor plenum with height of 76.2 cm (30-inch) or higher

had the most uniform airflow rates through the

perforated tiles and most uniform temperature across the

top of the racks. The findings also indicate that the

uniformity of the static pressure inside the plenum

increases with plenum height. The current results show

that data center plenums with height range of 76.2-91.4

cm (30-36 inch) are found to be the best option based on

the current thermo-fluid considerations however thermo-

economics analysis should be addressed in future work.

INTRODUCTION

Energy consumption is a critical factor in the design

and operation of data centers. The data center is the

. Environmental Protection

Agency (EPA) reported that data centers consumes at

least 1.5% of the total United States of America

electricity consumption [1] and this value is expected to

increase over time. This increase is motivated in part by

the technological advancement and increased demand

for data centers with more capacity and high

performance compute systems in addition to compaction

or consolidation of hardware to make better use of

physical floor space. Blade servers are great examples of

such compaction, and represent one of the fastest

growing enterprise server markets. The thermal

requirement of Information Technology (IT) equipment

is usually specified by temperature and airflow rate. This

requirement is met by delivering the proper amount of

airflow at the pre-specified temperature to the inlet side

of the IT equipment. Some studies [2-5] have discussed

the lack of energy efficiency in state of the art data

centers and showed that for every one watt of power

used by the IT equipment, an additional 0.5-1.0 Watt of

power is required by the cooling resources to maintain

proper inlet conditions for the same equipment i.e. an

additional 50%-100% is used to operate the cooling

resources. The over provisioning of the cooling systems

design can add millions of dollars to the annual thermal

management cost for large data centers. The capital cost

required to deploy cooling systems at this scale is also

considerable. Hence efficient cooling of data centers is

the way to minimizing power consumption, reducing the

harmful emissions to the environment and providing a

cost savings prospective.

Typical high density data centers are designed based

on hot/cold aisles layout. The conditioned air is drawn at

specific temperature and humidity into the IT systems by

the integrated fans or blowers. Hot air rejected from the

IT equipment is exhausted into the hot aisles and

recycled through the room which is known as room

return or through a ceiling plenum and in this case is

2

called ceiling return, back to the computer room air

handler (CRAH) units. Ideally, the hot return air streams

should return to the CRAC units without mixing with the

cool air from the supply. However, most data centers

have an open air return design (no ceiling plenum)

which results in mixing the servers air exhaust with

some of the conditioned air streams and reducing the

energy efficiency of the data center. The complicated air

flow distribution in the open environment makes thermal

management of such data centers challenging. The lack

of proper thermal environment inside a data center is a

serious threat not only to system reliability and

performance, but also warranties and service agreements

of systems manufacturers.

The design treatment of the heat rejected by the

computational, networking and storage resources into

the data center is specified simply by matching the

installed cooling capacity to the total data center design

power rating. However, this approach does not guarantee

may be formed. Hence a thermal requirement specified

by a combination of a temperature range and a minimum

airflow rate is required for reliable data center operation.

This specific requirement can be met by delivering

sufficient amount of cooled air to the inlet of each rack

at a temperature within the acceptable temperature range

.

The most common data centers are designed based

on raised-floor architecture and hence the focus of this

paper.

Fig. 1 Typical raised-floor data center configuration

Figure 1 depicts a simple data center configuration

with one CRAH pressurizing under-floor plenum

enclosure. The floor plenum should be free from any

significant blockages that can disturb airflow and the

static pressure inside the plenum. Plenum blockages may

be caused by network cable trays, power cables and

unused legacy cables from past installations. Overhead

racks trays design maybe a good solution to potential

plenum blockages.

A conventional raised-floor data center is typically

equipped with perforated tiles with fixed percent open

area and CRAH units with constant flow rates.

Perforated tiles are critical elements of the raised-floor

data centers as they are the channels by which the cooled

airflow is delivered to the data center IT systems. The

CRAH units pressurize the under-floor plenum and

maintain a positive static pressure differential between

the plenum pressure and above the raised-floor pressure.

The pressure differential is the driving force for the

cooled airflow across the perforated tiles with the

volumetric airflow being proportional to the percent

open area. Most CRAH unit manufacturers attempt to

thermally manage the data center environment based on

the CRAH unit return air temperature although there has

being a move to use the supply air temperature instead.

In either case, the set point temperature of each CRAH

unit is selected manually and is based on a qualitative

assessment of the data center environmental conditions

which makes provisioning data center cooling capacity

even more challenging.

DATA CENTER THERMAL MANAGEMENT

CRAH units and perforate tiles are two knobs that

are available to manage and fine tune the thermal

environment inside a data center and eliminate hot spots.

The data center thermal management is achieved by

selecting the proper CR individualized set

point temperatures and individualized airflow rate. Fine

tuning the localized airflow is achieved by selecting the

proper number, size and layout of the perforated tiles.

Selecting the number and size of the CRAH units are

typically performed quantitatively by matching them to

the total heat load generated inside the data center while

the data center layout of the CRAH units and the number

and layout of the perforated tiles are performed

qualitatively based on the IT equipment heat load

distribution. Caution should be taken since the IT system

utilization and consequently the heat load is

continuously fluctuating and as a result the cooling

capacity requirement is also fluctuating.

Off-the shelf data center perforated tiles are

available with a fixed percent open area ranging

Cold

Aisle

Server rack

CRAH

Unit

Perforated

Tile

Hot

Aisle

Plenum

Racks

Exhaust

3

typically between 25% and 56% although perforated

tiles with higher percent open area is also available in

the market. Facility engineers and operators are required

to physically move, relocate and replace perforated tiles

for local airflow adjustments and to match new cooling

requirements or prevent localized hot spots that may

result from insufficient cooled air circulation. The

relocation of perforated tiles is not a recommended long

term strategy since it introduces changes to the plenum

static pressure distribution, to the available airflow from

each individual perforated tile and may even disrupt the

overall data center thermal environment in addition to

being impractical and labor intensive. This issue

becomes particularly more pronounced when the data

center plenum height is less than 2-ft in height [6, 7].

Perforated tiles with automated dampers have been

introduced as a solution to the challenges that data

centers operators face in tuning the localized airflow to

match the continuous change in the heat load

distribution. These controllable perforated tiles tune the

airflow by adjusting the percent-open area between 0%

and 50% [6]. The adjustable dampers provide flexibility

to an existing standard perforated tile [7]. Monitoring

and controlling these channels will provide local thermal

management that could lead to energy savings [8, 9].

The increased fluctuation in the plenum static pressure

negatively affects the airflow uniformity through the

perforated tiles. This results in over provisioning of

cooling capacity to compensate for the airflow loss in

the affected areas and reduces the overall data center

energy efficiency. Published reports and studies are

available that highlight energy efficiency challenges and

opportunities in data centers [10-12]. New energy

efficiency metrics have being proposed to standardize

data center energy performance assessment based on

first principle and thermodynamic approach [13].

This paper presents the results of a numerical study

on the effect of a plenum height/size on the perforated

tile individual airflow rate and rack inlet temperature

distribution at 7-ft above the raised floor. This study

utilizes FloVent software, a commercially available

computational fluid dynamics (CFD) package. The

method takes into account the data center size, the

plenum dimensions, the CRAH set point temperature,

location and air flow rate, the perforated tiles number,

size and locations. The results are presented in terms of

the temperature, static pressure and volumetric flow rate

distribution. This paper provides insight into optimum

data center plenum sizing underscoring the power of the

numerical design approach. The importance of

predicting the performance of a given data center prior to

construction and facilitating changes to the layout of the

IT and cooling resources are evident.

NUMERICAL MODEL

The partial differential equations describing the

complex interactions between the dynamics of fluid flow

and heat transfer are highly coupled and non-linear

making them difficult to be solve analytically. Resolving

these equations numerically requires the use of various

approximation techniques such as finite volume, finite

element and finite difference. The partial differential

equations are initially discretized and transformed into

algebraic form to be solved iteratively until a pre-

specified convergence level is reached. The convergence

level is determined based on the variables residual

monitored relative to the problem scale at hand.

The governing equations describing the airflow are

continuity, momentum and energy equations.

(1)

Equation (1) is the continuity equation also known

as the conservation of mass equation. The variables

are the density, velocity and time, respectively.

This equation states that the mass entering a control

volume is equal to the mass leaving the control volume

plus the change of mass within the same control volume.

(2)

Equation (2) is the conservation of Momentum. The

variables ( ) are pressure, gravity and the viscous

shear tensor, respectively. This equation is derived from

the net

force is equal to mass times acceleration. The forces

exerted on the fluid volume in this case are pressure,

viscous and body forces. This equation is a vector

equation and therefore has a magnitude and a direction.

- (3)

4

Equation (3) is the conservation of Energy. The

variable ( ) is the total internal energy and the thermal

conductivity, respectively. This equation is derived from

the first law of thermodynamics and states that the net

exchange of energy across the control surfaces is equal

to the change between the initial and final state of the

system.

The numerical model of the data center requires

information on the actual and maximum heat load of the

IT equipment and the maximum available air flow and

capacity of the CRAH units. Caution should be

exercised to make sure that each CRAH capacity

in the numerical model does not exceed its actual

physical (rated) capacity at any given time. This can be

done by coding the CRAH unit maximum allowable

capacity into the numerical model and verifying that the

limit have not being exceeded by checking the final

results of the numerical model. The model, once

converged, solves for the temperature profile and airflow

distribution among other variables, and enables

optimization of the installed cooling capacity in the data

center.

STEADY STATE MODEL

The principal features of the mathematical model

presented in this paper are defined as a steady-state

three-dimensional model with negligible radiation effect.

The motion of the fluid layers within the space of the

data center is affected by fluctuating currents caused by

mixing, entrainment and circulation and therefore it is

best represented using the k-Epsilon (k-

model that calculates the kinetic energy and its

dissipation rate.The k-ε model solves for two variables:

k; the turbulent kinetic energy, and ε; the rate of

dissipation of kinetic energy. The k-epsilon model has

good convergence rate and relatively low memory

requirements. The numerical model is built taking into

account all of the data center details that affect air flow

and temperature distribution including plenum size,

load. The data center has a total area of 29 m2 (312 ft2)

with one row of ten industry standard server racks and

one perforated tile strategically located for each server

rack to provide the necessary cooling capacity. The

server racks are modeled as enclosures with open ends

based on the standard industry racks dimensions. The

model defines the servers as heat sources with their

corresponding airflow rates and each perforated tile as a

planar resistance to reflect the percent open area of 47%.

The selection of the 47% open perforated tiles is based

on experience and what is typically available as an off-

the shelf product. The maximum load of each rack

selected for this study is 10kW with a maximum heat

load of 100 kW for the overall data center. One standard

CRAH unit supplies the necessary cooling capacity for

the overall data center and is modeled to supply the

cooled air at a given temperature. In this model, the

CRAH unit supply temperature is set at 20 oC.

RESULTS AND DISCUSSION

This study presents a numerical analysis to predict

the impact of the changes in the plenum size on the

plenum static pressure, the airflow through each

perforated tile and temperature distribution above the

raised-floor. The objective of the analysis is to show that

the availability of adequate cooling resources to match

the heat load is not sufficient by itself to prevent hot

spots that may lead to server failures if and when non-

uniformity of airflow distribution reach a critical level

[4]. The numerical model is built based on the

assumption that there is no obstruction to the airflow

inside the plenum and that the plenum walls are

adiabatic i.e. the temperature at the exit of the perforated

tile is equal to the air temperature supplied by the CRAH

unit.

Figure 2 shows a sample temperature distribution

plots for various plenum sizes. The temperature shown

on the plots is the spatial temperature distribution at

213.4 cm (7-ft) above the raised floor. This location is

selected because it is typically the worst case scenario

for the inlet temperatures. The plots show hot spots

sizes 61 cm (24-inch) and smaller. A hot spot is defined

as the inlet air temperature at or above 30 oC. The inlet

air temperature distribution becomes more uniform as

the plenum size increases and the uniformity in

temperature becomes more pronounced for plenum sizes

91.4 cm (36-inch) and larger.

5

30.5 cm (12”) Plenum 76.2 cm (30”) Plenum

45.7 cm (18”) Plenum 91.4 cm (36”) Plenum

61.0 cm (24”) Plenum 152.4 cm (60”) Plenum

Fig. 2: Sample temperature plots at 213.4 cm (7-ft) above the raised floor for various plenum sizes.

The volumetric airflow rate from the each perforated

tile is plotted and shown in Fig. 3. Although the current

results are based on the model at hand and on the given

assumptions, they still valuable in terms of showing the

trend of how the volumetric air flow rates distribution

among the perforated tiles vary as a result of the plenum

size changes. There exists a plenum size range where the

volumetric airflow is more uniform and that is when the

plenum size is between 76.2-91.4 cm (30-36 inch).

Figure 4 displays the inlet temperature taken at

213.4 cm (7-ft) above the raised floor and at the inlet of

the top server in the rack. The plot shows that the inlet

temperature to some of the servers for 30.5 cm (12-inch)

and 45.7 cm (18-inch) plenum sizes approaches the

maximum operating threshold temperature set by the

manufacturer. Each system has a built-in thermal

threshold that may vary from one type of system to

another but typically is set to 36 oC at which the system

graceful shutdown starts. Data centers owners and

operators work hard to avoid this issue and prevent

g the critical

thermal threshold range and therefore it makes very

good sense to numerically evaluate the data center

environment numerically and before the actual systems

are in place. It is very difficult to mitigate this issue once

the data center is already in operation and the only

solution is to over provision the cooling resources and

manually redistribute/reselect the perforated tiles around

the affected systems. The current model results show

temperature slightly above 33 oC and approaching the

critical thermal threshold level.

Temp > 30oC

Servers’ Exhaust

Server Racks

CRAH Unit

Servers’ Inlet

T (oC)

6

Fig. 3: Individual perforated tiles volumetric flow rate

for various plenum sizes

Fig. 4: Inlet temperature at 7-ft (213.4 cm) above the

raised floor

The maximum room and inlet rack temperatures at 213.4

cm (7-ft) above the raised-floor as a function of the

plenum height are shown in Fig. 5. The plot shows both

temperatures to follow the same trend which indicates

the linear relationship between the inlet and the exhaust

temperatures. The plot also show that at plenum height

of 61 cm (24-inch) or smaller help the generation of hot

spots that may leads to over provisioning and reduces

the energy efficiency of the data center.

Fig. 5: Maximum room temperature and Inlet

temperature at 213.4 cm (7-ft) above the raised floor

versus plenum height

Tables 1 and 2 display sample numerical results

obtained from the CFD model. The results display the

inlet temperatures at 213.4 cm (7-ft) above the raised

floor as the raised-floor plenum size varied.

Perforated

Tile no.

Plenum Height

45.7 cm (18")

61.0 cm (24")

76.2 cm (30")

91.4 cm (36")

cm3/min

T , oC

cm3/min

T , oC

cm3/min

T , oC

cm3/min

T , oC

-5

16.0

33.8

22.5

28.0

31.9

22.9

40.8

21.3

-4

43.4

27.0

47.8

23.6

48.6

22.1

48.1

21.2

-3

59.2

20.7

54.3

21.7

52.1

21.5

49.7

21.0

-2

63.2

20.1

59.0

20.8

54.5

20.7

51.4

20.7

-1

59.0

20.0

57.2

20.4

53.6

20.8

50.7

20.6

1

59.0

20.0

57.2

20.4

53.6

20.8

50.7

20.6

2

63.2

20.1

59.0

20.8

54.5

20.7

51.4

20.7

3

59.2

20.8

54.3

21.7

52.1

21.5

49.7

21.0

4

43.4

27.0

47.8

23.6

48.6

22.1

48.1

21.2

5

16.0

33.8

22.5

28.0

31.9

22.9

40.9

21.3

Max. Room Temp.

44.6

38.8

35.6

34.8

Table 1: Airflow rate and temperature values for various plenum sizes at 213.4 cm (7-ft) above the raised floor

0

10

20

30

40

50

60

70

-5 -4 -3 -2 -1 1 2 3 4 5

Perforated tile flow rate, cm3/min

Perforated tile number

30.5 cm Plenum

45.7 cm Plenum

61.0 cm Plenum

76.2 cm Plenum

91.4 cm Plenum

106.7 cm Plenum

121.9 cm Plenum

137.2 cm Plenum

152.4 cm Plenum

15

20

25

30

35

40

-5 -4 -3 -2 -1 1 2 3 4 5

Temperature,oC

Perforated tile number

30.5 cm Plenum

45.7 cm Plenum

61.0 cm Plenum

76.2 cm Plenum

91.4 cm Plenum

106.7 cm Plenum

121.9 cm Plenum

137.2 cm Plenum

152.4 cm Plenum

0

5

10

15

20

25

30

35

40

45

50

020 40 60 80 100 120 140 160 180

Temperature, oC

Plenum height, cm

Maximum Room Temperature

Maximum Rack Inlet Temperature

7

Perforated

Tile no.

Plenum Height

106.7 cm (42")

121.9 cm (48")

137.2 cm (54")

152.4 cm (60")

cm3/min

T , oC

cm3/min

T , oC

cm3/min

T , oC

cm3/min

T , oC

-5

47.0

20.8

48.1

20.7

48.9

20.9

49.6

21.2

-4

47.8

20.7

47.6

20.6

47.7

20.7

47.9

20.7

-3

48.2

20.6

47.9

20.5

47.5

20.6

47.3

20.6

-2

48.9

20.4

48.3

20.4

47.9

20.5

47.6

20.6

-1

48.8

20.4

48.7

20.3

48.5

20.5

48.3

20.6

1

48.8

20.4

48.7

20.3

48.6

20.5

48.3

20.6

2

48.9

20.4

48.3

20.4

47.9

20.5

47.6

20.6

3

48.3

20.6

47.9

20.5

47.6

20.6

47.3

20.6

4

47.8

20.7

47.6

20.6

47.8

20.7

47.9

20.7

5

47.1

20.8

48.2

20.7

49.0

21.0

49.6

21.3

Max. Room Temp.

34.7

35.0

35.4

36.0

Table 2: Airflow rate and temperature values for various plenum sizes at 213.4 cm (7-ft) above the raised floor

It is clear from the tables that the bulk of the airflow

leaves the plenum through the middle perforated tiles for

plenum heights smaller than 76.4 cm (30-inch) and

tapers down towards both ends of the perforated tiles

row. The volumetric flow rate from each of these two

perforated tiles measured less than 50% of the average

air flow rate through any of the other individual

perforated tile. The variation between the perforated tiles

volumetric airflow rate ranges between 3-5% at the low-

end and 12-13% at the high-end of the total and

maximum volumetric flow rate of 481.4 cm3/min

(17,000 CFM) specified for the CRAH unit by the

manufacturer. The results also show the maximum

variation between any two perforated tiles is as high as

75% for 45.7 cm (18-inch) plenum size. This variation

between the low-end and the high-end volumetric flow

rates decreases to 20% for 91.4 cm (36-inch) plenum

size, to about 4% for the 42-inch (106.7 cm) plenum size

and to about 2.3% for the 121.9 cm (48-inch) plenum

size. Interestingly this variation goes slightly up to about

3% for the 137.2 cm (54-inch) plenum size and to close

to 5% for the 152.4 cm (60-inch) plenum size. The

temperature distribution at 213.4 cm (7-ft) above the

raised-floor plenum moves towards more uniformity for

plenum height of 76.2 cm (30-inch) and larger. The

results indicate that an optimum raised-floor plenum size

range exist for the constant airflow inlet condition

specified in this study.

SUMMARY AND CONCLUSION

A numerical model of a hypothetical density data

center is built to investigate the effect of plenum height

on the level of uniformity of the static pressure within

the plenum just below the perforated tiles. The static

pressure just below the perforated tiles directly affects

the perforated tiles airflow distribution and the inlet

airflow temperature to the server racks. The steady-state

model is created and the effect of the plenum height/size

on the airflow distribution and inlet server temperature is

analyzed. The steady state model shows that the rack

inlet temperature increases to an unacceptable level

when the plenum height is 61.0 cm (24-ft) or smaller.

This issue can be resolved by proper

redistribution/addition of the perforated tiles around the

affected servers. However this approach may affect other

areas negatively and disrupt the overall airflow

distribution and cooling capacity of the data center.

Numerical modeling and pre-analysis of the data center

environment is the best approach to predict and resolve

such issues before it happens. Furthermore, economic

assessment should be conducted when selecting the

plenum size for a new data center as larger plenums

require larger capital investment; the cost of the plenum

is not included in the current analysis. The following are

the specific key points summary of this study:

The uniformity of the static pressure inside the

plenum increases with plenum height. This uniformity

becomes more pronounced as the plenum height

increases beyond the 61 cm (24 inch).

8

The data center plenums with height range of 76.2-

91.4 cm (30-36 inch) are the best option based on the

current thermo-fluid considerations and for the specified

geometry, however thermo-economics analysis should

be addressed in future work.

High density data centers architecture and design

approach must be augmented with detailed

computational thermo-fluids analyses. Providing

sufficient cooling resources based on the total compute

heat load in the data center is not sufficient by itself, as

localized high power density will lead to localized hot

spots.

Modeling the thermo-fluids behavior in the data

center helps with uptime assessment, create policies and

procedures to avoid data center failure and enable

reliable data center operation.

Modifying the number or layout of the perforated

tile, adding or removing plenum obstructions such as

wires and cables will affect the static pressure uniformity

inside the plenum and may disrupts the airstream paths

however this behavior will be more pronounced for

smaller plenums heights. It is highly recommended that

all power and network cables, wires used inside the data

center are connected through rack-top cable trays.

One way of maintaining control over the static

pressure uniformity inside a plenum is to create dividers

(compartmentalization) and allow each CRAH unit to

feed air into a specific compartment within the plenum.

This approach provides a granular control of the static

pressure uniformity inside the plenum and allow for

tighter control of the volumetric airflow through the

perforated tiles. This method may take away the CRAH

units redundancy unless at least one side divider is

designed to be removable either automatically or

manually when the adjacent plenum section static

pressure goes below a certain threshold pressure value.

FUTURE WORK

Future work will include more detail evaluation and

analysis of an actual size data center. The current study

should be extended to investigate the effect of multiple

CRAH units and their zone of influence variation as the

plenum size changes. Direct versus indirect chilled air

delivery into the plenum effect on the static pressure and

consequently on the perforated tiles airflow rates should

be investigated.

REFERENCES

[1] US Department of Energy, Report to Congress on

Server and Data Center Energy Efficiency Public

Law 109-431 U.S. Environmental Protection

Agency ENERGY STAR Program , 2007.

[2]

Efficiency Optimization and GreeningCase Study

Buildings Journal, Elsevier, 2014.

[3] Patel, C.D., Bash, C.E., Sharma, R., Beitelmal,

A.H.,

center enabled by advanced flexible cooling

hermal Measurement

and Management Symposium, IEEE, 2005.

[4] -fluids

J. of

Distributed and parallel databases-special issue on

high density data centers, Vol. 21 No. 2-3, 227-238,

2007.

[5] Beitelmal, A.H., Patel, C.D., "A Steady-State

Model for the Design and Optimization of a

Centralized Cooling System," Int. J. of Energy

Research, Vol. 34 No. 14, 2010.

[6]

tates Patent No.

7,347,058, 2008.

[7] Beitelmal, A. H., McReynolds, A.A., Bash, C.E.,

Manipulating

United States Patent No. 8,639,651, 2014.

[8] controller for

7,902,966, 2011.

[9] Beitelmal, A.H., Wang, Z., Felix, C., Bash, C.,

Proceedings of the InterPACK2009 conference,

San Francisco, California, July 2009.

[10] Kaplan, J., Forrest, W., Kidler, N.,

McKinsey and Company, data center efficiency

report, 2008.

[11] Ebbers, M., Galea, A., Schaefer, M., Khiem, M.,

published by IBM, 2008.

9

[12]

distribution system's airflow performance for

cooling energy savings in high-density data

2014.

[13]

Buildings Journal, Elsevier, 2014.