Investigation of CRAH Bypass for Air-Cooled Data

Centers using Computational Fluid Dynamics

Vahid Ebrahimpour Ahmadi Hamza Salih Erden
Mechanical Engineering Department Informatics Institute
Istanbul Technical University Istanbul Technical University
Istanbul, Turkey Istanbul, Turkey
[email protected] [email protected]

Abstract—Computer room air handling (CRAH) bypass UA thermal conductance, kW/K

(BP) method utilizes fan-equipped perforated tiles to force or V velocity, m/s
induce room air through relatively lower flow resistance than W power, kW
that in CRAH units, as CRAH fans operate at lower speed.
Increasing the BP airflow rate reduces the total fan power and Greek symbols
increases the chiller power since the CRAH unit needs to Δ difference
provide colder air when a certain fraction of warm room air is ε thermal effectiveness, -
allowed to mix with cold CRAH airflow. Simplified airflow ρ mass density, kg/m3
network and thermodynamic modeling tools addressed this Φ CRAH-to-rack air flow ratio, -
optimization problem and indicated cooling energy savings
potential via CRAH BP method for enclosed aisle (EA) data Subscripts
centers. These tools assume uniform air temperature at the bp bypass
server inlets, even though recent experimental and fan fan
computational studies indicate temperature non-uniformities
c cooling
at higher fractions of BP flow. This study addresses energy
implications of the uniform temperature assumption of the
ch chiller
reduced order modeling tools on the cooling energy saving chw chilled water
potential of EA data centers by using computational fluid cr CRAH
dynamics (CFD) modeling. Moreover, existing literature does i inlet
not thoroughly investigate the application of CRAH BP in the max maximum
open aisle (OA) data centers, which requires CFD models to o outlet
resolve complex airflow patterns due to recirculation of hot air r rack
into the cold aisle and leakage of cold air into the hot aisle. This t tile
study utilizes CFD to investigate both induced and forced
CRAH BP methods in a representative quadrant of a 1MW I. INTRODUCTION
data center with and without aisle containment.
In recent years, with the development of data centers, the
Keywords—CRAH bypass method, air-cooled data centers, industry devoted many efforts to improve the energy
computational fluid dynamics efficiency of data centers. Power Usage Effectiveness (PUE)
is the most widely used energy efficiency metric for data
NOMENCLATURE center infrastructure, which is the ratio of total facility
energy consumption to that of the information technology
Acronyms (IT) equipment [1]. Despite considerable efforts regarding
BP bypass the data center energy efficiency, reports show more than
CA cold aisle 55% of studied cases had a PUE greater than 1.8 [2]. As a
CFD computational fluid dynamics result of this survey, most data centers experience inefficient
CRAH computer room air handling/handler cooling systems and require effective air and thermal
EA enclosed aisle management.
fBP forced CRAH bypass Server racks are typically located back to back and front
FNM flow network model to front in air-cooled data centers to form hot and cold
HA hot aisle aisles. Schmidt et al. [3] analyzed different parameters
HX heat exchanger including ceiling and raised floor height, as well as the
iBP induced CRAH bypass direction of airflow, which influence the cooling efficiency
IT information technology of air-cooled data centers. They concluded that mixing of
OA open aisle cold and hot air is the main problem in raised floor open
QDC quadrant data center aisle (OA) air-cooled data centers. The recirculation
TDM thermodynamics model between cold and hot aisle leads to noticeable temperature
non-uniformity at server inlets. To keep the maximum
Latin symbols temperature at server inlets below the 27°C, which is the
COP coefficient of performance, - recommended upper limit by the American Society of
P pressure, Pa Heating Refrigerating and Air Conditioning Engineers
K loss coefficient, - (ASHRAE) [4], chillers need to work inefficiently at lower
NTU number of transfer units, - temperatures.
PUE power usage effectiveness One practical way to reduce the temperature non-
T temperature, °C uniformity is applying enclosed aisle (EA). Reduced
This research is supported by the Scientific and Technological temperature non-uniformity allows the cooling system to
Research Council of Turkey (TÜBİTAK) through Award Number
118M238.
operate at a higher temperature possibly taking advantage of positive plenum pressure when the CRAH fans run at lower
more free cooling hours. However, CRAH units in raised speeds. Therefore, floor leakage remains intact. In contrast
floor EA data centers have to supply the entire airflow for to the iBP fans that supply the entire rack airflow, fBP fans
servers, as well as wasted airflow through the leakage paths supply just a fraction of the rack airflow (Fig. 1b).. Erden et
and cable cutouts. Hence, an important part of cooling al. [9] estimated five percentage points more reduction in
infrastructure energy consumption is due to the CRAH fans cooling power by iBP compared to fBP using flow network
[5]. modeling (FNM) for predicting fan power and
CRAH bypass (BP) method, recommended by Khalifa thermodynamic model (TDM) for estimating the chiller.
and Demetriou [6][7][8], decreases the CRAH air flow and Researchers studied CRAH BP method in various details
provides the complementary fraction of the tile airflow through different modeling tools and experiments. FNM is
either by forcing or inducing warm air from the hot aisle useful for predicting airflow rate and fan power especially
into the plenum using BP fans. This approach leads to for EA data centers [9][10]. Erden et al. experimentally
significant reduction in the total fan power since the validated FNM in air-cooled data centers with and without
pressure BP fans need to overcome is typically an order of BP with decent accuracy [11][12]. FNM predictions of the
magnitude less than that CRAH fans do. However, reduced airflow rate and fan power are within less than 10% of
fan power does not guarantee optimum operation since the experimental measurements [12]. Even though Erden et al.
cooling system needs to operate at a lower temperature [12] pointed out higher temperature non-uniformity at
inefficiently to keep the rack inlet air temperatures below higher BP fractions, they used well-mixed temperature
27°C. Hence, there is a trade-off between the fan power and assumption in the thermodynamic modeling (TDM) of the
cooling power. data center for optimization purposes in various previous
Fig. 1 shows the schematics of two configurations of studies [9][10].
CRAH BP method, induced BP (iBP) and forced BP (fBP). Computational fluid dynamics (CFD) is a powerful tool
In the iBP configuration, iBP fans are located just below the that allows us to test the well-mixed temperature assumption
perforated tiles in the cold aisle (CA). At reduced CRAH for a range of BP fractions and quantify the errors associated
flow rates, iBP fans compensate the reduced air flow by with this approach. A preliminary assessment by Erden
inducing room air through openings across the raised floor. shows that the temperature non-uniformities at larger BP
These openings can be inherent leakage ports (e.g., cable fractions may lead to different results [13]. Unlike the
cut-outs, tile seams, etc.) and intentional openings (e.g., accomplished studies, which consider CRAH BP method
perforated tiles in the hot aisle). The pressure in the plenum just for EA data center, CFD also helps us to determine the
decreases at lower CRAH flow rates and even becomes complex airflow pattern and feasibility of CRAH BP
lower than that of the room especially when CRAH flow method for OA data centers. This study focuses on both
rate is lower than rack air flow rate. Negative pressure in the induced and forced CRAH BP methods in a representative
plenum causes the room air to move from the room into the quadrant of a 1MW data center with and without aisle
plenum eliminating the leakage airflow (Fig. 1a). containment. The following sections discuss both modeling
approaches (TDM and CFD) before introducing the
optimization procedure and results.
II. METHODOLOGY
Fig. 2 provides simplified schematics of the FNM and
TDM, which are based on commercially available modeling
tools, AFT Fathom [14] and TRNSYS [15], respectively.
The flow network of the baseline data center configuration
excludes the air flow path with the solid green line and both
iBP and fBP fans. CRAH BP configurations, iBP and fBP,
have the air flow path with solid green line activated to
allow the CRAH BP air flow in addition to the
corresponding BP fans (i.e. iBP and fBP). In the FNM, the
flow resistance of the CRAHs, racks, perforated tiles and
leakage paths are the dominant flow resistances. The
pressure loss due to various resistances can be computed by,
1
P  K V 2 , (1)
2
where the K is the specific loss coefficient, ρ is the air
density and V the air velocity. Assuming zero-static pressure
for the room and relevant fan characteristics [10], the
Fig. 1. Representative view of (a) induced and (b) forced BP obtained pressure and flow distribution on the flow network
configurations for air-cooled data centers lead to the estimation of fan power.

On the other hand, the fBP method consists of fBP fans

underneath perforated tiles in the hot aisle (HA) preferably
adjacent to the CRAH units. These fans maintain the
 Fig. 3. Top view of the QDC layout for both (a) iBP and (b) fBP
configurations.
Fig. 2. Schematics of TDM and FNM.
Abdelmaksoud et al. [19] experimentally validated data
Considering the fan flow rate and power use estimated center CFD models at room level wıth uniform 4-inch grid
by the FNM and setting the rack inlet temperature at the size. They noticed that finer grid did not improve the results
redline temperature, TDM determines well-mixed air unless some of the details about the perforated surfaces are
temperatures at key locations (e.g., plenum, tile, rack exit, included. The higher numerical diffusion caused by the
CRAH inlet/exit, etc.). The ε-NTU relations provides the coarse grid compensates the less diffusive entrainment due
thermal effectiveness, ε, of the CRAH heat exchangers (HX) to the coarse representation of perforated surfaces. Hence,
[16], where NTU is the ratio of the thermal conductance UA the computational region for the QDC consists of uniform
to the minimum of the fluid heat capacity ratios. At reduced hexahedral mesh with a size of 2 inches (0.0508 m) for the
CRAH air flow rates, UA scales with the 0.8th power of the plenum and 4 inches (0.1016 m) for the room to maintain
flow rate according to the change of Nusselt Number for y+ values within recommended limits [20]. The numerical
turbulent internal flow with Reynolds Number. Hence, HX simulation has been executed by using the commercially
model determines the chilled water temperature (Tchw) to be available, CFD software, Fluent 18.1 [21], by applying the
supplied by the chiller model. The chiller model computes standard k-ε turbulence model, standard wall function, and
the Coefficient of Performance (COP) and chiller power full buoyancy effects.
based on a normalized performance map at various chilled QDC consists of three halves and one whole CRAH
water setpoint temperatures [17][18]. units, indicating 10 CRAH units for the entire data center.
CFD modeling allows us to solve the mass (continuity), Constant downward velocity inlet boundary condition in the
momentum and energy equations to determine the plenum represents CRAH units’ exit condition with a
temperature and flow field in the air space of the data center. uniform temperature. QDC domain has 20 racks in two rows
The layout of the 22 x 38 ft (6.71 x 11.58 m) quadrant of a with uniform IT load distribution housing 60 blade-type
generic data center (QDC) for both configurations servers. The recirculation model allows reduced-order
investigated in this paper is shown in Fig. 3 (top view). In modeling of the server without the need for meshing the
addition, the ceiling height above the raised floor and under- rack interior [20]. This model couples inlet and exit surfaces
floor plenum height are 144 inches (3.66 m) and 36 inches of each server, and adds generated IT load heat to the inlet
(0.91 m), respectively. The dimensions of each floor tile are airflow to determine the exit boundary condition. The
24 x 24 inches (0.61 x 0.61 m). turbulence length for inlet boundaries is considered 7% of
the hydraulic diameter of the opening [22], with typical
values for turbulence intensity of 10% for servers [23] and
50% for CRAH units [8].
The cold aisle in the QDC is 3-tile wide and 10-tile long.
The rest of the floor consists of two different openings to
allow airflow between the room and plenum. First, leakage
tiles behind the racks representing cable cutouts; second,
bypass tiles adjacent to CRAH units. Based on Erden et al.
[12] all openings either in the cold aisle or hot aisle have a
flow resistance coefficient of K=21. A 4-inch (0.1016 m)
thick porous volume assumes these openings with a porosity
of 25% and the inertial resistance coefficient along the
vertical flow direction. The location of CRAH BP fans
depend on whether iBP or fBP approach is in use. The fans
in the iBP are installed underneath the perforated tiles in the
cold aisle (Fig. 3a) and in the fBP, only underneath the
perforated tiles adjacent to CRAH units in the hot aisle (Fig.
3b). However, for both configurations, the fan boundary
condition causes a constant pressure rise across the fan Table 1 lists the assumptions for the 1 MW IT load data
boundary surface, which is parallel to perforated tiles and center. The equally-distributed IT load among the racks
placed 12 inches (0.3 m) away from the tiles inside the causes rack airflow temperature to increase 14.6°C. The
plenum. entire data center has 120 perforated tiles in the cold aisle.
According to the CFD study by Demetriou and Khalifa, These tiles house 120 BP fans in iBP configuration (Fig.
the BP tiles adjacent to CRAH units provide more time for 3a). CRAH BP configurations have 40 perforated tiles in the
mixing and having more uniform temperatures at the cold HA next to CRAH units in addition to the leakage paths
aisle and rack inlets [8]. Thus, BP fans in fBP configuration behind the racks. 40 BP fans operate underneath these tiles
as shown in Fig. 3b are intentionally underneath perforated for fBP configuration (Fig. 3b). Leakage path flow area and
tiles next to CRAH units. Similarly, iBP has these resistance is equivalent to 40 perforated tiles and causes
purposely-placed perforated tiles next to CRAH units to approximately 25% of the CRAH air flow rate to leak into
induce room air into the plenum in addition to the inherent the room in the baseline configuration (i.e., Φ=1.33). Key
cable cutouts. Also, all of the cases have been studied both flow characteristics such as the resistances (CRAH, racks
with and without cold aisle containment (EA and OA). and tiles) and fan curves (CRAH, server and BP) are based
The parametric CFD analyses provide temperature and on vendor data and in line with the assumptions in the
flow fields for the QDC for various CRAH-to-rack airflow previous optimization studies [10]. TDM model uses
ratios (Φ) for iBP and fBP with EA and OA data center. Fig. thermal conductance UA for the cross-flow HX of a CRAH
4 shows the synopsis of the procedure for the TDM-based unit extracted from vendor catalog data. Chiller cooling
optimization of CRAH BP method for only EA data center capacity is 1500 kW, has a COP of 6.0 and the condenser
[13] and CFD-based optimization for both EA and OA data side receives cooling water at a temperature of 35°C.
centers.
TABLE I. SUMMARY OF MODEL ASSUMPTIONS
IT Power 1000 KW
# of Racks 80 -
°
Rack Air Temperature Rise 14.6 C
# of CRAH Units 10 -
Single CRAH Rated Flow Rate 8.02 m3/s
Single CRAH HX UA at the Rated Flow Rate 16.0 kW/K
CRAH Pressure Loss Factor, Kcr 141 -
Server Pressure Loss Factor, Kr 205 -
Perforated Tile Pressure Loss Factor, Kt 21 -
# of Perforated Tile in the Cold Aisle 120 -
# of Fans for iBP Configuration 120 -
# of Fans for fBP Configuration 40 -
Chiller Rated Cooling Capacity 1500 kW
Design COP of the Chiller 6.0 -
°
Condenser Water Temperature 35 C

III. RESULTS AND DISCUSSION

Fig. 5 compares BP fan power predictions by CFD and
FNM for both induced and forced EA configurations. In this
figure, for a range of Φ values, the relative standard
deviation in iBP and fBP fan power Wbp between the CFD
Fig. 4. Optimization procedure using TDM and CFD for various Φ values. and FNM is 14.7% and 1.5%, respectively. Considering the
total fan power, Wf , on the secondary vertical axis,
As the first step, to obtain zero-static pressure inside the
aisle containment or rack inlets, CFD simulation provides the comparison of CFD and TDM approaches results in a
required uniform static pressure increase across the BP fans negligible error. This confirms the decent capability of FNM
(ΔPbp). Given the pressure and airflow, the operating points as a prediction tool for air-cooled data centers by Erden et
on the BP fan and efficiency curves provide the power use al. [11].
by the BP fans (Wfan,bp). The obtained Wfan,bp is an energy
source input in CFD causing a temperature rise across the
tiles equipped with BP fans. Ultimately, CFD provides the
CRAH air supply temperature (Tcr,o) which keeps the
maximum rack inlet air temperature (Tr,i-max) below the
redline (27°C). Finally, according to the mass-weighted inlet
and exit air temperatures at the CRAH unit, the chiller
power, Wch, can be computed by utilizing the CRAH and
chiller components of TDM (Fig. 2). The resultant cooling
power is

Wc = Wch + Wfan,cr + Wfan,bp , (2)

where Wfan,bp is already based on CFD and CRAH fan

power (Wfan,cr) is the same as that obtained by the FNM. Fig. 5. Comparison of BP and total fan power estimations by TDM and
CFD.
 Fig. 6 and 7 shows the cooling power (Wc) and the evaporator temperatures compared to the EA configuration.
associated chilled water temperatures (Tchw) based on TDM Fig. 8 shows Wc and Tchw based on TDM and CFD for OA
and CFD, for both iBP and fBP configurations, respectively. configuration. Both iBP and fBP exhibit a gradual decrease
The dashed red lines represent Wc by the TDM assuming in cooling power especially down to Φ=1.1, which stems
well-mixed temperature and pressure assumptions at all from the rapid reduction in CRAH fan power. The cooling
nodes for EA data center. The solid red lines are Wc for power difference between the iBP and fBP configurations in
CFD. Also, solid and dashed blue lines indicate the OA data centers is not as meaningful as it is in EA data
corresponding Tchw based on CFD and TDM, respectively. centers. However, it is important to note that fBP method
As the amount of BP flow increases (Φ approaches 0.7), the uses only one-third of the BP fans iBP uses in this study.
difference between TDM and CFD model increases. Unlike
the fBP configuration, for iBP configuration the deviation is
remarkable. Reduced mixing and total plenum flow rate due
to the elimination of leakage flow in the iBP configuration
can be a possible explanation for higher temperature non-
uniformity at lower Φ values. It is noteworthy to mention
that optimum Φ values by TDM are typically lower than
that by CFD. In CFD the optimum operation point for both
iBP and fBP configurations are about 0.9≤Φ≤1; however, it
is around 0.8≤Φ≤0.9 in TDM.

Fig. 8. Comparison of cooling power for both iBP and fBP in open aisle
configuration using CFD.

Considering the optimum Φ value around 0.9 for all

possible cases in this study, increasing the CRAH BP
airflow beyond this value does not have an advantage.
Around Φ=0.9 for iBP configuration, Wc decreases about
21% and 17% respectively, for EA and OA. Similarly, for
fBP configuration, Wc reduces by 20% and 18%. These
results indicate energy saving potential of CRAH BP
Fig. 6. Comparison of cooling power for iBP configuration by TDM and
method for not only EA but also OA data centers.
CFD. The BP fan model in this CFD study assumes a uniform
pressure rise across all BP fans ignoring the complex flow
pattern downstream that can further enhance mixing. More
advanced fan representations including swirl functionality
and testing system design variations are also possible in
CFD. Future work also needs to include the economic
considerations of possible CRAH BP configuration
investment for both EA and OA air-cooled data centers as a
retrofit. The sensitivity of these results to the efficiency of
the existing cooling infrastructure and other design
parameters will also help the industry to have a better
understanding on the feasibility of the CRAH BP method.
IV. CONCLUSION
In order to evaluate the energy-saving potential of
CRAH BP method, past studies assumed well-mixed
Fig. 7. Comparison of cooling power for fBP configuration by TDM and pressure and temperature in FNM and TDM. According to
CFD.
the recent experimental investigations, despite acceptable
performance of FNM in predicting the airflow rate and fan
Since the variation of the Φ values changes the entire
power, it does not adequately consider effects of
airflow field in the room and plenum for OA data centers,
temperature non-uniformities on the optimization results.
obtaining generalized conclusions regarding the application
The current research applies CFD method for different
of CRAH BP method is more complicated compared to that
configurations in a range of CRAH BP fractions to estimate
for EA data centers. Since the fans operate against same
BP fan power and temperature non-uniformities. The
flow resistance, the BP fan power for both iBP and fBP in
obtained 14.7% and 1.5% relative standard deviation
OA cooling system is the same as that of the EA system.
between FNM and CFD respectively for iBP and fBP
However, due to the higher temperature non-uniformities in
represents FNM as a decent tool in predicting BP fan power.
OA data center, the cooling system operates at lower
For Φ≥1.0 optimization results with TDM has a good Transactions on Components, Packaging and Manufacturing
Technology, vol. 7, Issue 12, pp. 1981-1989, 2017.
agreement with CFD results in the EA data center. In
[11] Erden, H.S., Koz, M., Yildirim, M.T., Khalifa, H.E., "Experimental
contrast, for Φ ≤ 1.0, CFD results significantly deviate from investigation of CRAH bypass for enclosed aisle data centers," in
TDM, which assumes well-mixed temperature in modeling. Proc. IEEE ITherm, pp. 1293-1299, Las Vegas, NV, USA, May/Jun.
The deviation is more considerable for iBP configuration 2016.
rather than fBP. Based on CFD results, the optimum Φ [12] Erden, H.S., Koz, M., Yildirim, M.T., Khalifa, H.E., “Experimental
value is approximately 0.9 for all configurations. Around Demonstration and Flow Network Model Verification of Induced
CRAH Bypass for Cooling Optimization of Enclosed-Aisle Data
this value, the Wc related to the iBP decreases by 21% and Centers,” IEEE Transactions on Components, Packaging and
17% for EA and OA, respectively; likewise, 20% and 18% Manufacturing Technology, vol. 7, Issue 11, pp. 1795-1803, 2017
for fBP. [13] Erden, H.S., “Investigation of Induced CRAH Bypass for Air-Cooled
Data Centers Using Computational Fluid Dynamics,” in Proc. IEEE
REFERENCES SoftCOM, Split, pp. 1-6, 2017.
[1] Belady, C., Rawson, A., Pfleuger, D., and Cader, T., “Green Grid [14] FATHOM 9, Incompressible Pipe Flow Analysis & System Modeling
Data Center Power Efficiency Metrics: PUE and DCiE,” The Green Software, Appl. Flow Technol., Colorado Springs, CO, USA.
Grid, White Paper No. 6, 2008. [15] Klein S., “TRNSYS—A transient system simulation program,” Solar
[2] Stansberry, M., and Kudritzki, J., “Uptime Institute 2012 Data Center Energy Lab., Univ. Wisconsin-Madison, Madison, WI, USA, Eng.
Industry Survey,” Uptime Institute, New York, 2013. Experim. Station Rep. 38-13, 2000.
[3] Schmidt RR, Cruz EE, Iyengar MK. “Challenges of data center [16] Incropera, F. and DeWitt D. et al., Fundamentals of Heat and Mass
thermal management,” IBM Journal of Research and Development, Transfer, John Wiley & Sons, 2007.
vol.49, Issue 4.5, pp. 709–23, 2005. [17] Braun, J.E., “Performance and control characteristics of large central
[4] ASHRAE, Thermal Guidelines for Data Processing Environments, cooling systems,” ASHRAE Transactions 93(1), 1830–52, 1987.
4th Edition, American Society of Heating, Refrigerating and Air- [18] Demetriou, D.W., H.E. Khalifa, M. Iyengar, and R. Schmidt,
Conditioning Engineers, Inc., Atlanta, 2015. “Development and experimental validation of a thermohydraulic
[5] Salim, M., and Tozer, R., “Data Centers’ Energy Auditing and model for data centers,” HVAC&R Research vol. 17, Issue 4, pp.
Benchmarking-Progress Update,” ASHRAE Trans., Vol. 116 Issue 1, 540–555, 2011.
pp. 109–117, 2010. [19] Abdelmaksoud, W. A., Dang, T. Q., Khalifa, H. E., and Schmidt, R.
[6] Khalifa H.E., Demetriou D.W., “Energy Optimization of Air-Cooled R., “Improved Computational Fluid Dynamics Model for Open-Aisle
Data Centers,” ASME Journal of Thermal Sciences and Engineering Air-Cooled Data Center Simulations,” ASME Journal of Electronic
Applications, vol. 2, Issue 4, 041005-041005-13. Packaging, vol. 135, Issue 3, p. 030901, 2013.
doi:10.1115/1.4003427, 2011. [20] Versteeg, H.K. and W. Malalasekera, An Introduction to
[7] Khalifa, H.E., Demetriou, D.W., “Enclosed Aisle Data Center Computational Fluid Dynamics: The Finite Volume Method, Pearson
Cooling System,” U.S. Provisional Patent Application No Education Limited, 2007.
61/367,931, 2010. [21] ANSYS, Fluent 18.1 User Guide. Knowledge Resources: Online
[8] Demetriou, D.W., Khalifa, H.E., “Optimization of Enclosed Aisle Documentation.
Data Centers Using Bypass Recirculation,” Journal of Electronic [22] ANSYS, Fluent 18.1 User Guide, Chapter 6: Cell Zone and Boundary
Packaging, vol. 134, Issue 2, doi:10.1115/1.4005907, 2012. Conditions.
[9] Erden, H.S., Yildirim, M. T., Koz, M. and H. E. Khalifa, “Energy [23] Abdelmaksoud W. A., “Experimental and Numerical Investigations
assessment of CRAH bypass for enclosed aisle data centers,” in Proc. of the Thermal Environment in Air-cooled Data Centers,” PhD
IEEE ITherm, pp. 433–439, Las Vegas, NV, USA, May/Jun. 2016. Dissertation, Syracuse University, NY, 2012.
[10] Erden, H.S., Koz, M., Yildirim, M.T., Khalifa, H.E., “Optimization of
Enclosed Aisle Data Centers with Induced CRAH Bypass”, IEEE