IWSE Reg

Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

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

net/publication/264982820

Simulation of a temperature adaptive control strategy for an IWSE economizer


in a data center

Article  in  Applied Energy · December 2014


DOI: 10.1016/j.apenergy.2014.07.072

CITATIONS READS

40 488

4 authors, including:

Baptiste Durand-Estebe Cédric Le Bot


Nobatek Ecole Nationale Supérieure de Chimie, de Biologie et de Physique (ENSCBP)
7 PUBLICATIONS   84 CITATIONS    35 PUBLICATIONS   536 CITATIONS   

SEE PROFILE SEE PROFILE

Eric Arquis
Ecole Nationale Supérieure de Chimie, de Biologie et de Physique (ENSCBP)
115 PUBLICATIONS   1,491 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Baptiste Durand-Estebe on 14 September 2018.

The user has requested enhancement of the downloaded file.


Applied Energy 134 (2014) 45–56

Contents lists available at ScienceDirect

Applied Energy
journal homepage: www.elsevier.com/locate/apenergy

Simulation of a temperature adaptive control strategy for an IWSE


economizer in a data center
Baptiste Durand-Estebe a,b,⇑, Cédric Le Bot a,1, Jean Nicolas Mancos b, Eric Arquis a
a
Univ. Bordeaux, I2M, UMR 5295, F-33400 Talence, France
b
Cap-Ingelec, Allée des Palanques, 33127 St Jean d’Illac, France

h i g h l i g h t s

 The CFD model of a 32 kW data center is presented.


 A Proper Orthogonal Decomposition (POD) method is described.
 A cooling plant model is presented.
 A consumption analysis in function of the air temperature is carried out.
 A new regulation strategy for the cooling plant is proposed.

a r t i c l e i n f o a b s t r a c t

Article history: Nowadays, with the constant evolution of Information Technology (IT) equipments, the energy consump-
Received 17 December 2013 tion of data center over the world becomes a major concern. In 2011 the ASHRAE Technical committee 9.9
Received in revised form 16 July 2014 (TC9.9) issued important guidelines concerning server temperature and hygrometric environment to help
Accepted 20 July 2014
engineer in the design of cooling solutions. While raising the temperature may be a source of heat pump
Available online 23 August 2014
energy savings, it induces an increase in the Computer Room Air Handling (CRAH) unit energy require-
ment, lowering the benefits. Hence optimal temperature cooling set point must be found to maximise
Keywords:
the efficiency of the cooling plant. To test various chiller control strategy a ‘‘full scale’’ model is proposed.
Data center
CFD
A 32 kW data center is considered, cooled by a centrifugal heat pump linked to a wet cooling tower. An
Free cooling Integrated Water Side Economizer (IWSE) is added to minimize the energy consumption the regulation
Regulation and the chilled air production is simulated with the software TRNSYS. The temperature field in the server
POD room is calculated with the CFD code Thetis. To create a link between the 2 simulation environments, a
Cooling plant Reduced Order Model (ROM) using Proper Orthogonal Decomposition (POD) is program with MATLAB.
Finally this numerical model is used to investigate the effect of server room temperature increase on
the cooling plant energy consumption. A new Temperature Adaptive Control Strategy (TACS) that mini-
mizes the energy need is proposed and tested.
Ó 2014 Elsevier Ltd. All rights reserved.

1. Introduction researches are carried out to lower the overall energy needs.
According to Patterson [3] 3 sources of energy consumption can
Data centers are large infrastructures housing hundreds of IT be identified: the IT part gathering all the servers and computing
servers working 24/24, 365 days/year [1]. With the constantly units, the Cooling part that includes the machines involved in the
increasing material performance, their consumptions are quickly chilled air management, and the power supply system. The present
raising. According to prediction, the overall energy absorption study is focused on the relations between the servers and the cool-
should reach 100 billion kW h around the year 2016 [2]. Therefore, ing plant, thus the last pole of consumption won’t be taken into
for environmental and economic considerations, numbers of account.
The paper presents the energy optimisation of a 32 kW data
center by numerical simulation. The server room is composed of
⇑ Corresponding author at: Univ. Bordeaux, I2M, UMR 5295, F-33400 Talence, 32 servers mounted in rack and arranged in cold aisle/hot aisle
France. Tel.: +33 (0)5 4000 62 11. Under-floor – Locally Supply/Locally Return U–LS/LR according to
E-mail address: [email protected] (B. Durand-Estebe).
1
[1]. The cooling plant is composed of a Heat Pump (HP), a cooling
Tel.: +33 (0)5 40 00 61 92.

http://dx.doi.org/10.1016/j.apenergy.2014.07.072
0306-2619/Ó 2014 Elsevier Ltd. All rights reserved.
46 B. Durand-Estebe et al. / Applied Energy 134 (2014) 45–56

Nomenclature

CTKE LES model constant


Model variables a LES model constant
Pit server power dissipation (kW) T filtered temperature (°C)
0
Mit server fan flow rate (m3/h) T temperature standard deviation (°C)
MCO CRAC output flow rate (m3/h) k thermal conductivity (W/mK)
TCO CRAC output air temperature (°C) kt turbulent thermal conductivity (W/mK)
TCI CRAC input air temperature (°C) Cp air heat capacity (J/kg K)
TII server inlet temperature (°C) Prsg sub-grid Prandtl number
TIImax inlet temperature of the hotter server (°C)
Tloop CRAC inlet water temperature (°C) POD snapshot variables
Tchhot CRAC outlet water temperature (°C) Ns number of snapshots
TchHX IWSE heat exchanger outlet water temperature (°C) x number of scalar node in a snapshot
Tchi chiller evaporator inlet water temperature (°C) Tp matrix gathering snapshots
Tcho chiller evaporator outlet temperature (°C) Tm mean matrix
Tcoo chiller condenser outlet temperature (°C) s spatial deviation
Tcoi chiller condenser inlet temperature (°C) R correlation matrix
Tsump cooling tower sump water temperature (°C) k eigenvalues
t eigenvectors
Mechanic equations / eigenmode
u filtered air velocity (m/s) C POD coefficient
q air density (kg/m3) Enrj eigenmode energy
p filtered air pressure (Pa) E max3600 prediction maximum error on scalar nodes (°C)
l dynamic viscosity (Pa s) RMS3600 prediction RMS error on scalar nodes (°C)
k permeability (m2) E max 32 prediction maximum error on server temperature (°C)
lt turbulent dynamic viscosity RMS32 prediction RMS error on server temperature (°C)
CM LES model constant HAOC Hourly Average Operation Cost (kW)
CS,1 LES model constant

tower and a Computer Room Air Handling (CRAH) unit. It also inte- dramatic inflation of CRAH fan energy consumption. Thus for a spe-
grates a Water Side Economizer to enhance its energy efficiency cific cooling plant, the best compromise between air supply tem-
with an appropriate controller. perature and air flow rate has to be found to minimize the
The CRAH units maintain the IT material in very particular air overall energy absorption. This set point depends on many param-
conditions. In 2011, the ASHRAE TC9 representing IT equipment eters including: cooling plant design, material and Air Distribution
manufacturers created several environmental envelops bounded System (ADS) performance.
by temperature and hygrometry [4]. These recommendations aim This optimal trade off between air temperature and flow rate
to give more flexibility to engineers to conceive innovative cooling will also be affected by the integration of an Economizer in the
solutions for data centers while insuring maximum security and cooling plant. Several studies pointed that using an economizer
reliability. The 4 environmental classes are plotted on the psycho- (airside or waterside) and raising the supply air temperature may
metric chart Fig. 1 lead to good results [5,10]. In fact, depending on the climate and
In the study, the modelled data center belongs to the A1 class on the data center air temperature set point, the whole system
with the most restrictive temperature and hygrometry conditions. can achieve free cooling or partial free cooling operation during
While raising the chilled air temperature can be a way to reduce hundreds of hours a year. This ‘‘free’’ cooling energy will constantly
the overall energy consumption by lowering the Heat Pump (HP)
energy needs, this solution has to be handled with care. The ASH-
RAE guideline [4] indicates that most of the recent IT servers are
sensitive to their inlet temperature. A hotter cooling airflow
induces an increase in current leakage at a ship level. In recent
blade server, the Fan Speed Control Algorithm (FSCA) reacts and
raises the server’s fan speed to maintain the material under accept-
able temperature. This behaviour leads to an undesirable increase
in the server energy consumption. Several studies have been car-
ried out to estimate the impact of raising the temperature in a
recent data center, and the effect on the servers [5–8]. Hence low-
ering the HP energy requirement by only increasing the supplied
air temperature is not necessarily an efficient way to save energy.
Moreover in order to maintain the air at the intake of all rack
under a maximum temperature, a specific CRAH air flow rate has
to be found. The ‘‘physic based simple model’’ introduced by
Demetriou [9] shows that in an open aisle data center, hotter tem-
peratures do not always lead to energy savings. As the supply air
temperature rises, the required air flow rate to maintain the IT
equipment under a maximum temperature increases leading to a Fig. 1. ASHRAE environmental classes for data centers.
B. Durand-Estebe et al. / Applied Energy 134 (2014) 45–56 47

modify the balance between the consumption of the cooling pro- in cold aisle/ hot aisle. The server room is cooled by a cooling plant
duction handled by the HP and of the air distribution achieved composed of a chiller, a 4 cells wet cooling tower, a CRAC unit, a
by the CRAH fans. water/water heat exchanger (IWSE), 2 circulation pumps and 3
The aim of the study is to find through simulation, an adaptive con- mixing valves. A programmed controller and two PID regulators
trol strategies that will constantly optimize the energy performance maintain the different temperature and flow rate set points.
by adjusting the air temperature and the air flow rate set points.
But when it comes to the economizer selection, a choice has to
be made between air-side and water-side systems [10]. In their 2.1. The room layout
study, Bulut and Aktacir [11] studied the free cooling potential of
an air-side economizer used in an office cooling plant, located in The server room is a classic U–LS/LR configuration. The chosen
the city of Istambul. They determine that rising the temperature geometry corresponds to a part of the data center study by [1].
up to 24 °C, allows the cooling plant to work under free-cooling The dimensions are reported Figs. 2 and 3 and Table 1:
condition during 89% of the year. However, the hygrometric and The characteristics of the IT servers are the same as described
temperature requirements for office environment are not as by Cho et al. [1]. A simple model has been chosen compared to
demanding as server room environment. Thus, Dai et al. [12] stud- the one described in [8]. In this paper the power and the fan flow
ied the risk of using an air-side economizer in a datacenter cooling rate are considered constant. The aim is to present an innovative
plant and proposed to reduce it by using a prognostics-based way of modelling a whole data center and to establish links
approach to prevent the failure risks. Finally, according to Myer between a classic the server room environment condition and
and Sorell [10], air-side economizers are generally more efficient the cooling plant regulation. In further study the interaction
then water-side economizers. However the second class of system between high density blade server and cooling plant will be
presents several advantages when it comes to data center rehabil- studied.
itation such as small footprint and low investment cost. Moreover The CRAH flow rate is regulated by a PID controller to ensure a
unlike the air-side economizer [13], no detailed control strategy security condition while minimizing the fan speed. This will be dis-
have been proposed for water-side economizer. Therefore, the cussed in the corresponding chapter. Finally to ensure mass con-
IWSE free cooling system is selected for the study, and its model- servation, the flow rate absorbed by the two ventilation grid
ling is used to improve its regulation strategy. located on the hot aisle at the ceiling are equal and compensate
However the simulation of a server room and its cooling plant the inlet flow rate blown at the centre in the cold aisle.
over a year presents numbers of issues. In order to propose an
overall simulation of a data center, several tools have to be used
and linked together. The outdoor climate, the cooling plant, the 2.2. The cooling plant and its regulation strategy
Integrated Water Side Economizer (IWSE) and the controller will
be simulated on the TRNSYS software. But as the plant controller
The full scheme of the cooling plant with a schema of the server
requires the instantaneous temperature of all the IT equipments,
room is presented on Fig. 4. The dotted line in the server room
a temperature field must be computed inside the room. The CFD
mark out the boundaries of the CFD simulation:
code Thetis developed by the I2 M laboratory handled the velocity
The 1st aim of a data center is to ensure a safe work environ-
and temperature field calculation inside the server room. Unfortu-
ment for the IT equipments. Thus to make sure that every IT server
nately in order to simulate a full year (8760 h), TRNSYS requires a
operates under a given temperature, a PID regulator adjusts the
CFD calculation each 0.5 h corresponding to its simulation time
flow rate MCO. The numerical model computes the servers’ intake
step, while the CFD code requires at least 24 h to complete a calcu-
air temperature and sends the hottest result TIIMAX to the room reg-
lation. In this condition it is not conceivable to run the two pro-
ulator. The controller varies the CRAH fan speed to maintain this
grams in parallel. Thus to estimate server room air condition a
temperature under a given set-point based on the ASHRAE recom-
Reduced Order Model (ROM) using Propper Orthogonal Decompo-
mendations [4]. The chosen TIIMAX set-point for this study corre-
sition (POD) with the snapshot technique [9] is used to quickly
sponds to the maximum allowable temperature 32 °C.
interpolate CFD results obtained from Thetis. This technique is
implemented in TRNSYS using the Matlab software. It provides
an interesting environment to test new control strategy and to
estimate their impact on energy savings over a year.
The first part of this article presents the 32 kW data center that
will be modelled for this study. In a second part the numerical
methods used to create the simulation environment are presented.
Then the effects of rising the supply air temperature on the energy
consumption are investigated. Finally a thermally aware control
strategy adapted to IWSE is proposed.
This work aims to use the cold/hot aisle configuration intro-
duced by Cho [1] in the 32 kW server room, and to link it to the
IWSE model proposed by Mackay [18]. The POD Snapshot tech-
nique presented by Demetriou [9] is used, and a new way of mod-
elling full scale datacenter is proposed. At last, the obtained model
is used to propose an original and innovating type of adaptive reg-
ulation called TACS.

2. The data center layout

The infrastructure simulated in this paper is a small computer


room for a total constant power of 32 kW. The racks are arranged Fig. 2. Top view of server room configuration.
48 B. Durand-Estebe et al. / Applied Energy 134 (2014) 45–56

flow rate is equally shared between the four cells and each cell’s
fan is constantly loaded at 70% of its capacity.
The last control component is the IWSE regulator appearing on
Fig. 4. The aim of this device is to optimize the installation in func-
tion of several temperatures. Its algorithm is derived from the
paper [10] and is presented on Table 2.

2.3. The cooling plant equipments

The Heat Pump is the central component of the cooling installa-


tion. Its rated capacity is 44.6 kW with a rated Coefficient Of Per-
formance (COP) of 4.5. Its other characteristics such as Part load
ratio (PLR) and COP ratio are similar to the ones described in
[14]. The evaporator side is connected to the CRAH unit. The vari-
able speed pump P1 characteristics of this circuit are taken from
the Wilo products [15].
Fig. 3. Side view of server room configuration. The CRAC unit that blows the cool air in the server room is com-
posed of a water/air heat exchanger and a variable speed fan. The
characteristics of the heat exchanger are taken from the Alfa Laval
catalogue [16], the computed heat transfer coefficient is
Table 1
Variables definition.
UA = 31707 kJ/h K. The fan rated flow rate is 8453 kg/h for a rated
power of 5.31 kW, these data are derived from the TRANE cata-
W Room Width 5.0 m logue [17]. It is considered that the flow rate can reach 120% of
H Room Height 3.5 m
RL Rack Length 1.0 m
the rated characteristic for a short period of time.
Rw Rack Width 0.8 m The heat pump evacuates the calories with a cooling tower con-
Rh Rack Height 2.0 m nected on the condenser side. A 0.835 kW single speed pump P2
Haw Hot aisle width 0.5 m [15] drives the water in the circuit. The cooling tower contains 4
Caw Cold aisle width 1.0 m
cells connected to a single 2.4 m3 sump. The temperature of the
AL Aisle Length 3.2 m
d Aisle distance to the wall 0.35 m replacement water entering the sump is 20 °C. Each cell has a
ITh IT server height 0.5 m 1.12 kW variable speed fan with a maximum flow rate of
MIT IT server fan air flow rate 160 m3/h 4000 m3/h.
Pit IT server power release 1 kW Finally a 0.9 constant efficiency water/water plate heat exchan-
ger is added to the cooling plant according to the schemes [10,18].
It allows the energy extracted from the server room to flow from
the evaporator side to the condenser side without the use of the
The second PID regulator adjusts the water flow rate in the heat pump. Two by-pass valves driven by the IWSE controller are
evaporator loop to maintain the temperature TCO at the required added on each side of the heat exchanger. The one on the evapora-
set-point. In this first control strategy TCO is constant over a year. tor side V2 regulates the water temperatures TCHI returning to the
In order to keep reasonable flow rates in the evaporator loop, a chiller. The one on the condenser side V3 by-pass the heat exchan-
DT of 7 °C is observed between the water temperature entering the ger when non free-cooling operations to avoid additional pressure
heat exchanger Tloop and the blown air set point temperature TCO. drop.
The heat pump set-point temperature varies from 9 °C to 10 °C Finally a third loop-by-pass valve V1 also driven by the IWSE
depending on TCO. For Tloop temperatures greater than 10 °C, a tem- controller is added to the evaporator. As the Heat pump maximal
pering valve V1 partially mixes the returning hot water to the evaporator temperature is TCHO = 11 °C, the valve V1 allows loop
chiller outlet cold water to get the right set point. water temperature to increase by by-passing a part of the return-
On the condenser side, the cooling tower is fed by a single speed ing fluid. Doing so, the system can increase the number of hours
pump P2 that provides a constant 7000 kg/h water flow rate. The of free-cooling or partial free-cooling.

Fig. 4. Cooling plant schematic.


B. Durand-Estebe et al. / Applied Energy 134 (2014) 45–56 49

Table 2
IWSE Economizer mode selection process.

The blackened devices are switched off or unused


Free-cooling (WSE) Partial free-cooling (IWSE) Normal
TchHX 6 Tloop Tchhot > Tsump AND NOT(WSE) Tsump > Tchhot
 The heat pump is turned off, The heat  The heat pump is turned on  The heat pump is turned on
load is evacuated through the cooling  The cooling tower works at 70% of its maximal  The cooling tower works at 70% of its maximal capacity
tower capacity  The 2 by-pass valves on the evaporator V2 and on the con-
 The cooling tower works at 70% of its  The 2 by-pass valves on the evaporator V2 and on denser side V3 are open. The heat exchanger is short cir-
maximal capacity the condenser side V3 are closed. The water of the cuited to reduce the pressure loss in the circuits
 The by-pass valve on the condenser 2 circuits flow through the heat exchanger.  The loop by-pass valve V1 is regulated, mixing chilled
side V3 is closed the water fully flows  The loop by-pass valve V1 is regulated, mixing water from the HP and returning water to maintain Tloop
through the IWSE heat exchanger chilled water from the Heat Pump and returning at the desired set-point
 The evaporator by-pass valve V2 is water to maintain Tloop at the desired set-point
regulated to maintain Tloop equal to
the desired set-point
 The loop by-pass valve V1 is closed

3. Flow and heat transfer equations ity, adding the turbulent viscosity lt. Thus the Eq. (2) can be written
under the following form:
Previous study [8,19] regarding the physical phenomena hap-    
@u
pening in a server room has shown the following results: q þ r  u  u ¼ qg  rp þ r  2ðl þ lt ÞS ð3Þ
@t
(a) The flow is fully turbulent in the whole data server room. In this paper both the Smagorinsky and the Turbulent Kinetic
(b) Heat transfers are mainly ruled by mixed convection. Energy (TKE) models are used to model the sub-grid viscosity
[20]. The turbulent viscosity is then written under the following
The filtered Navier–Stokes equations are used to compute the form:
air flow, combined to a Large Eddy Simulation (LES) model to take  a=2 h i1a
into account the turbulent effects. The air flow is considered lt =q ¼ C M D1þa 2Sij Sij ðqsm Þ1=2 ð4Þ
incompressible, the thermal properties are assumed constant (con-
ductivity, heat capacity, viscosity). The filtered momentum equa- qsm is the sub grid turbulent kinetic energy written
tions that are solved for the whole domain are written as follows qsm  1=2ðui Þ0 ðui Þ0 and ðui Þ0 represents the smallest resolved.
[20]: CM being a constant derived from the Smagorinsky and the TKE
models. It’s written under the following form:
ru¼0 ð1Þ
    l C M ¼ ðC s;1 Þ2a ðC TKE Þ1a ð5Þ
@u
q þruu ¼ qg  rp þ r  2lS  sij þ u ð2Þ Cs,1 = 0.173 is a constant from the Smagorinsky model and
@t k
CTKE = 0.2 is a constant from the TKE model. Finally a is set equal
In the Eqs. (1) and (2) u expresses the filtered velocity. On the left to 0.5 in order to ensure a good balance between the two sub-grid
side of the momentum equation @u @t
represents the unsteady acceler- models.
ation and r  u  u the convective acceleration. The right side of Eq. The filtered energy equation is written [21]:
(2) represents the sum of all the forces applied to the fluid. The first !
@T  Y
term is the gravity term and qB = 1.17 kg/m3 is considered constant. qC p þ u  rT ¼ r  krT  ð6Þ
The second right hand-side term rp is the pressure gradient term @t
with p the filtered pressure field. The parameter k is the permeabil- Q
ity, used here to model solid zones. Finally the last term Similarly to Eq. (2), the sub-grid stress tensor can be
r  ð2lS  sij Þ is composed of the viscosity constraint and the sub- expressed as a turbulent conductivity kt and Eq. (6) becomes:
!
grid stress tensor sij. The filtered rate-of-strain tensor S is written @T  
S ¼ 1=2ðru þ rT uÞ. In the LES approach, the energy dissipated by qC p þ u  rT ¼ r  ðk þ kt ÞrT ð7Þ
@t
the small scale is taken into account by increasing the local viscos-
50 B. Durand-Estebe et al. / Applied Energy 134 (2014) 45–56

The turbulent conductivity kt is modeled using a sub-grid Pra- Concerning the thermal boundary conditions, the 6 walls are
ndtl number Prsg method [21] with the equation: considered adiabatic. The temperature on the floor at the vent tile
location is equal to TCO.
lt C p
kt ¼ ð8Þ
Prsg
6. The POD snapshot method

CFD calculations are very time consuming. The resolution of one


4. Numerical methods simulation of the previously described model takes about 20 h
with 64 processors sharing the domain. Yet the simulation of the
The finite volume CFD code Thetis was used for all the simula- cooling plant behaviour would ask at least one simulation for every
tions that are presented in this paper. The numerical code is devel- 17 520 required time step to complete a calculation over a full
oped by the I2 M laboratory. It is based on a One Fluid model and year. Thus coupling the CFD code Thetis with the software TRNSYS
implicit finite volumes method. for a full calculation isn’t worth considering. However an access to
The equations are solved on a staggered grid to compute the a strong calculator (Jade CINES) allows carrying out a reasonable
coupling between pressure and velocity. The filtered Navier–Stokes number of simulations. Then a statistics-based reduced-order
equations are solved on the whole domain, a Brinkman term is model using Proper Orthogonal Decomposition (POD) with the
added to distinguish the fluid zone (air) for which a high value of Snapshot technique is developed on the software MATLAB and
permeability is given (practically k = 10+20 m2) and the solid zone linked to TRNSYS. Based on an initial set of data (the Snapshots)
(datacenter structure) corresponding to a low value of permeabil- this method permits to estimate a temperature field at the intake
ity (k = 1020 m2), so that in this zone, the velocity field is naturally of a server in few milliseconds and with a very small error.
set to u ¼ 0. Thus, air flow cannot penetrate the solid structure and The two parameters that would be controlled in the server
must bypass it. room are the intake temperature TCO and the cooling air input flow
The velocity – pressure coupling is ensured by a pressure pre- rate MCO. The information needed to control the server room are
diction/correction algorithm, based on the incompressibility con- the input temperature of each IT equipment, and the air tempera-
straint. Such a method is described by Goda [22]. The space ture returning to the CRAH unit TCI. to create the snapshots, 16 sim-
discretization is carried out by a theta scheme, and an Euler ulations are carried out for 4 different fan capacity (50%, 75%, 100%,
scheme is used for the time discretization. The resulting linear sys- 125%) and 4 different temperatures (16 °C, 18.67 °C, 21.33 °C,
tem is solved, using a BICGS solver [23] of the Hypre library, with 24 °C). For each simulation, the temperature field at the centre of
Jacobi preconditioning. The projection step is ensured by a HYPRE- the cold aisle corresponding to the server input and the mean tem-
BICG solver and a multigrid preconditioning. perature of the returning air to the CRAH TCI are extracted as shown
The energy equation is written in temperature. A hybrid scheme on Fig. 5.
is used for the space discretization. A first order time discretization
is also used. The linear system is solved by a BICG solver and ILU 6.1. The construction of the POD model
preconditioning.
The residues are fixed at 108 for the flow equations and 1016 The reduced order model is constructed according to the
for the energy equation to provide an accurate solution. method described by Demetriou [9] and Bui-Thanh et al. [26].
These numerical methods have been used and validated for the The following equations (from 9 to 14) are derived from their work.
simulation of the physical phenomena involved in a data center The 16 temperature fields are gathered in a single matrix Tp e MNs,x
server room. The following works by Nicolas et al. [24,25] present with Ns the number of snapshots and x the number of scalar nodes
the validation of the CFD code on a benchmark solution for a three- in the temperature field. The matrix Tm e Mx representing the
dimensional mixed convection flow. mean of the 16 temperature fields is created according to the Eq.
(10):
5. The server room model PNs
i¼1 Tpi;j
Tmj ¼ 8j 2 ½1; x ð9Þ
Ns
The server room is modelled on the software ECOTECT where
simple geometry and grid mesh can be easily defined. Further s e MNs,x is calculated following Eq. (11) representing the spatial
details on the model and the numerical method are reported in deviation of a snapshot from the mean field
the paper [8].
si;j ¼ Tpi;j  Tmj ; 8i 2 ½1; Ns; 8j 2 ½1; x ð10Þ
The square grid size is 60  60  60. The servers are modeled by
an open square enclosure and scalar nodes are placed on each par- The correlation matrix R e MNs,Ns is created according to Eq.
tition. A mesh refinement is used close to the air/structure inter- (12):
face to diminish the wall ‘‘diffusion’’.
Each server dissipates a constant heat power of 1 kW equally 1 X x
Ri;j ¼ si;k  sk;j 8i; j 2 ½1; Ns ð11Þ
distributed over its volume. At each scalar node inside the open Ns k¼1
enclosure, a source term is added to the energy equation to simu-
The eigenvalues k 2 M Ns and eigenvector t e MNs,Ns of the matrix
late the hardware energy dissipation.
R are computed and arranged such as k1 > k2 > ::: > kNs . The
The server fans are simulated by converting an air flow of
eigenvector are used to compute the eigenmode of the POD
160 m3/h into an air velocity and applying it at each scalar node
/ e MNs,x following the Eq. (12):
located at the outlet of the server.
The ventilation grids at the floor and at the ceiling are delimited X
x

by scalar nodes. The airflows are part of the boundary conditions /i;j ¼ si;k  tk;j 8i 2 ½1; Ns; j 2 ½1; x ð12Þ
k¼1
and the air speed is calculated from the mesh definition. This
way it guaranties the mass conservation of the system and the flow Finally the POD coefficients C e MNs,Ns are computed from the
incompressibility assumption. eigenmodes and the eigenvectors according to Eq. (13):
B. Durand-Estebe et al. / Applied Energy 134 (2014) 45–56 51

Fig. 6. Energy content of POD modes.

Fig. 5. Cold aisle temperature field [K]. ture, the reconstructed snapshots shows a E max32 = 0.12 °C and a
RMS32 = 0.05 °C. These results indicate that the POD method is able
Px to accurately reconstruct the snapshots with only 8 modes.
si;k  /k;j
C i;j ¼ Pxk¼1 8i; j 2 ½1; Ns ð13Þ In this study the POD model is used to reproduce the room tem-
k¼1 /i;k  /k;j perature field behaviour without the need to carry out large CFD
Using the previously computed Matrix each Snapshot can be calculation. Indeed, as the temperature field will be regulated by
reconstructed from the mean matrix Tm, the eigenmodes / and a PID controller, the fan speed and the air temperature TCI will vary
he POD coefficient C following the Eq. (14): linearly during the simulation. When the PID command does not
correspond to one of the 16 snapshots, the program uses the
X
Ns
POD model as an interpolation tools. It recreates 8 eigenmodes
T i;j ¼ Tmj þ C i;k  /k;j 8i 2 ½1; Ns; 8j 2 ½1; x ð14Þ
by performing a half spline half linear interpolation between the
k¼1
eigenmodes of the closest Snapshots. An interpolated temperature
field is re-constructed following Eq. (6). Hence the program is able
6.2. Investigation of the model accuracy to compute any temperature field for TCI ranging from 16 °C to
24 °C and fan flow rate ranging from 50% to 125% of the nominal
As described earlier, eigenvalues and eigenmodes have been value.
arranged by eigenvalue according to a decreasing order. Hence Several test where performed to validate the ability of the POD
the first POD eigenmode / computed with the first eigenvector t model to perform temperature field’s interpolations. When com-
is the most energetic. According to Demetriou [9], the energy paring the interpolated results to CFD simulation, it revealed a rel-
Enrji e MNs of the eigenmodes can be computed following Eq. (15): atively high maximum error of E max3600 = 3.05 °C but a small RMS
error of RMS3600 = 0.64 °C on the full temperature field. Consider-
ki
Enrji ¼ PNs 8i 2 ½1; Ns ð15Þ ing the server intake temperature, the maximum temperature
i¼1 ki error was E max32 = 0.85 °C and the RMS error was RMS32 = 0.32 °C.
The results of this calculation are plotted Fig. 6. These errors are not insignificant, however they do not comes from
The energy calculation indicates that 99.98% of the energy is the POD method as it showed good results when re-constructing
carried by the first 8 eigenmodes. Hence the re-constructed snap- an original snapshots. Moreover the accuracy of the interpolation
shot with these 8 eigenmodes should provide enough accuracy. method has been demonstrated in [9,27]. The errors probably
Further calculations with 16 eigenmodes would take longer to come from the lack of snapshots (only 16) and from the great tem-
solve and would not bring lots of additional information. To verify perature field variations between them. However it appears that
the accuracy of the reconstruction, four different errors formula are the model slightly overestimates the temperature, hence it should
defined. The maximum error on the 3600 scalar nodes between an penalize the energy gain and the obtained results won’t be overes-
original Snapshot and a reconstruction is E max3600. The Root Mean timated. Moreover, as the aim of this paper is to introduce a new
Square error on the temperature field prediction defined by Deme- way of modelling a full scale data center and to test a new type
triou [9], can be written: of regulation, these errors will not significantly impact the results.
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 Xx
RMS3600 ¼ ðTsi  Tri Þ 8i 2 ½1; x ð16Þ 7. The TRNSYS numerical model
x i¼1

where Ts e Mx contains the temperature of the original Snapshot TRNSYS is a very powerful environment made up of two parts.
and Tr e Mx contains the temperature of the reconstructed Snap- The kernel that read input files, solves the system, determines con-
shot. Similarly E max32 is defined representing the maximum error vergence and plots system variables. The components library con-
between the original server temperature and the reconstructed tains the weather data processors, the basic HVAC and the
ones. Finally RMS32 is computed, defining the RMS error in the ser- regulator units (www.trnsys.com). For this simulation, an addi-
ver intake temperature between the original simulation and the tional IWSE controller has been programmed and added to the sys-
reconstructed Snapshot. tem. It controls the by-passing valves and the HP as described in
The reconstruction test of a snapshot shows maximum error in Table 2. Furthermore, it switches off the HP when the evaporator
the 3600 scalar nodes temperature field of E max3600 = 0.48 °C is flow rate is null. The references of the main components (Type)
RMS3600 = 0.09 °C. When calculating the 32 servers input tempera- are given Table 3. Few additional details are given in this study
52 B. Durand-Estebe et al. / Applied Energy 134 (2014) 45–56

Table 3 Concerning the hygrometry, the highest humidity ratio is


TRNSYS types used in the cooling plant model. recorded for the month of December (83%) and the lowest for the
Components names TRNSYS month of July (73%). The average humidity ratio over a year is 78%.
type In order to determine the optimum temperature set-point for
TRNSYS – Matlab link (running the POD model) Type155 the cooling air in the server room, 5 simulations are carried out with
TMY2 weather data processor Type15 TCO ranging from 16 °C to 24 °C. However, as described in chapter
Centrifugal water cooled chiller (heat pump) Type666 2.3, a temperature drop of 7 °C is maintained between TCO and Tloop,
Cooling tower: user-supplied performance coefficients Type51
Pump – inlet flow rate known – power coefficients Type743
thus while raising the air supply temperature, the temperature of
Heat exchanger with constant effectiveness (water/water Type91 the cooling water in the loop on the evaporator side Tloop will also
IWSE) increase. The chilled water temperature set point of the heat pump
Cross flow heat exchanger: cross flow (water/air CRAC unit) Type5 Tchi is set as high as possible with a maximum of 10 °C to reach the
Variable speed fan/blower (CRAC unit fan) Type662
lowest energy absorption. Thus for TCO = 16 °C, Tchi = 9 °C, and for
IWSE controller Type808
PID controller Type 23 temperature greater or equal to 17 °C, Tchi = 10 °C. Concerning Tloop,
Controlled flow diverter Type11f temperatures higher than 10 °C are obtained by mixing the chilled
Tee piece Type11 h water with a part of the hot water returning from the IWSE heat
exchanger thanks to V1.
In the room, the intake air temperature of the IT equipments is
as the mathematical descriptions of the models are fully available calculated with the POD model. For each time iteration in the sim-
in the reference book [28,29]. ulation, the room PID controller determines the hottest server.
For the water cooled chiller model, the capacity ratio, the COP Then the regulator adjusts the chilled air flow rate MCO to maintain
ratio and the Part Load ratio (PLR) files are extracted from the ref- this hottest equipment at 32 °C, corresponding to the maximum
erence [14]. These data shows very little difference with the ones allowable temperature according to ASHRAE [4]. This choice of reg-
given by TRNSYS. In the same way the two coefficient required ulation is quite ‘‘aggressive’’ as it makes the servers run at the
by the cooling tower model are also the one given in [14]. For highest allowed temperature, but it enables the best energy sav-
the single and variable speed pump, the same model is used; the ings by setting the CRAC fan flow rate at the minimum required.
only difference lies in the flow rate set-point (constant or variable).
The pumps energy absorption is modelled by a normalized first
order curve as a function of the flow rate. This curve and the 8.1. Analysis of the results for TCO = 16 °C
corresponding coefficients are derived from the data of an
IPL-WILO-VEROLINE. The CRAH unit’s fan energy absorption is cal- The simulation is carried out over a year. Fig. 7 plots the utiliza-
culated the same way but with a simple 3rd order curve. Concern- tion ratio of the 3 modes WSE, IWSE and normal. Their values are
ing the PID regulation, the model proposed by TRNSYS is based on displayed on the left vertical axis. The curve plots the monthly
the algorithm given by Aström et al. [27,30]. This model uses a sin- mean temperature; the values are displayed on the right vertical
gle constant KPID for the three terms (Proportional, Integral and axis.
Derivative), plus 2 constants TINTand TDER to weight the influence In winter (from the 12/21 to the 03/21) time, the cooling towers
of the Integral and the Derivative Terms. As it is very difficult to sump temperature often falls below Tloop = Tchi = 9 °C and the heat
obtain a transfer function of the system, several tests have been pump can be switched off about 37% of the season to perform free
performed to define the optimal values of the constants for the cooling operation (WSE). During the remaining 63% the system
two PIDs. The values were selected to avoid oscillations, and to works on partial free cooling (IWSE) and the heat pump is very
get a good compromise between a fast response and a reasonable lightly loaded with an average of 10% of rated power required.
overshoot. However, and as reminded by the TRNSYS manual The first 15 h of the simulation are performed on forced ‘‘normal’’
[28], given the large number of PID algorithm, these optimal set- operation. It explains the few percentage of normal mode appear-
tings will probably vary when transferred to a real-life controller. ing on Fig. 6, but it does not alter the results. The mid-season gath-
Finally concerning the weather data, the hourly file comes from ers the spring season (from the 03/22 to the 06/21) and the autumn
METEONORM for the city of Bordeaux. The city is located in the (from the 09/21 to the 20/12). During this period, due to tempera-
south west of France, north hemisphere (44°500 600 North 0°340 4600 ture raise, partial free cooling is used 93% of the times whereas
west) It is created from meteorological measurement carried out free-cooling only represents 7%. Finally in summer time (from
over the past 30 years, and includes dry bulb air temperature, the 06/22 to the 09/20), the system works on partial-free cooling
hygrometry, solar radiations, etc. Further analysis of the tempera- 95% of the time. During the remaining 5%, the cooling tower sump
ture and hygrometry will be given in the next chapters.

8. Simulation results for various cold loop temperature

This chapter presents the results of 5 simulations of the data


center behaviour over a full ‘‘typical’’ year in the city of Bordeaux,
the weather file comes from the software METEONORM. The cli-
mate is Mediterranean with no extreme temperature either in
summer, or in winter. The weather file indicates that the coldest
temperature happens during the month of January (5.6 °C) which
is also the coldest month of the year with a mean temperature of
5.81 °C. During the summer, the hottest temperature is reached
during the month of August (32 °C). However the hotter month is
July with a mean temperature of 20.36 °C. Finally, over a year,
the mean dry bulb temperature is around 13 °C. Fig. 7. IWSE mode repartition during the year.
B. Durand-Estebe et al. / Applied Energy 134 (2014) 45–56 53

Fig. 8. Energy consumption repartition over a year for TCO = 16 °C.

temperature is greater than the temperature of the water exiting Fig. 9. Energy and IWSE mode evolution for various TCO.
the CRAC. The Economizer is no longer useful and the IWSE heat
exchanger is by-passed.
The components energy absorption over a year is plotted Fig. 8. replaced by WSE operations and the heat pump can be switched
As expected, the overall energy consumption strongly varies off during more hours. It can be noticed that this evolution is
with the seasons. At the beginning of the simulation, the room nearly a linear function of TCO. However the power consumption
PID regulator determines the right flow rate MCO to maintain the does not behave the same way. Great energy saving can be
server at the designed temperature. Hence the CRAC fan speed achieved between 16 °C and 22 °C but the gain between 22 °C
does not vary and its power consumption stands still along the and 24 °C is very small (0.8%). This phenomenon can be explained
year. The second PID regulator on the evaporator side behaves with Fig. 10.
the same way and sets the evaporator pump speed P1 to maintain The energy consumption of the cooling tower stands still as the
TCO = 16 °C. Concerning the cooling towers, the IWSE controller sets control maintains the command at 70% of its rated capacity over
their cells fans flow rate at 70% of their rated capacity. Finally the the year. Concerning the water pumps, their energy absorption
pump on the condenser side P2 is a single speed. Hence the only increase can be neglected as they represent a very little part of
equipment that does not have a fixed energy requirement is the the overall energy needs. However, great energy savings are
Heat Pump. It varies according to the seasons and depending on achieved through the heat pump. Thanks to the increase of the
the waterside economizer use. During the winter the load on the number of hours when the system perform free and partial free
heat pump is very low, the average temperature of water returning cooling operations, the annual average load decrease from 27% to
to the CRAC is 15 °C leading to a very low temperature drop 15%. The energy gain on the heat pump is quiet important and
between evaporator input and output (6 °C). These conditions lead its consumption is reduced by 78%. But these savings are counter-
to a lower mean COP of 5.09, but the light load on the heat pump balanced by a CRAH fan power increase. In fact, in order to main-
makes it the less expensive season of the year. During spring the tain the hotter server at the maximum allowed temperature, the
heat pump load rises, as the cooling tower is less efficient. The server room PID regulator increases the flowrate MCO by 14%. But
mean load is 26% for an average COP of 5.09. The behaviour during the relationship between the fan flow rate and the energy needs
the season of autumn is very similar. During summer the tower is is non linear [5] and the consumption increases by 43%. Thus the
less efficient and the IWSE system has to be switched off for several advantage of raising the blowing air temperature is limited and it
hours to perform Normal operation. The average load on the heat is likely that a simulation with TCO = 26 °C would have given an
pump is 44% and the COP slightly rises up to 5.94. energy consumption greater than the simulation with TCO = 24 °C.
The following paragraph presents several simulation results
with higher TCO temperatures. The aim is to allow the economizer
to increase the number of hours spent on WSE and IWSE opera-
tions. However while the heat pump consumption is supposed to
decrease, the CRAC fan consumption is expected to rise to maintain
the hottest server at the given set-point. It will generate an
increase in energy consumption that will lower the benefits of
more WSE and IWSE hours.

8.2. Analysis of the energy consumption variations for various TCO

Five simulations have been carried out for TCO = 16 °C, 18 °C,
20 °C, 22 °C, 24 °C. Rising the cooled air temperature, allows the
system to increase the Tloop temperature. It will get more advantage
from the Economizer as the limit condition to switch from the WSE
to the IWSE mode TchHX 6 Tloop will be harder to reach. Similarly the
use of the normal mode will be reduced as the CRAC outlet water
temperature Tchhot will increase. The results of the simulated
energy consumptions are presented Fig. 9.
As expected, the overall energy consumption goes down as the
percentage of WSE hours goes up. The IWSE operations are Fig. 10. Energy consumption repartition for various TCO.
54 B. Durand-Estebe et al. / Applied Energy 134 (2014) 45–56

Table 4
Cooling plant energy absorption per hour [kW] displayed by month for various Tco.

However such simulation was impossible to carry out as the CRAH temperature TCO. In fact as the heat pump energy goes down with
fan reached its limit speed. a reduction of the load when increasing TCO, it is compensate by a
This series of simulation highlights two main results. First, the fan power increase, giving HAOCIWSE = 6.2 kW for every simulation.
optimum blowing temperature for this system with this specific Table 4 presents the monthly energy consumption of the cooling
material and a regulation over the server maximum allowed tem- plant for the 5 simulations. For each month different shade of grey
perature is TCO = 24 °C. Raising the temperature set-point from are used to emphasize the difference in energy needs.
16 °C to 24 °C induces a gain of 10% on the overall energy con- As it can be seen on Table 4, the system is more economic in
sumption. But any modification in the chilled air production or winter time with TCO = 20 °C. This temperature is the best compro-
on the control strategy will probably move this optimal set-point. mised between the number of WSE operations and the HAOCWSE.
Second, the CRAC fan energy consumption tends to minimize the During mid-season and summer time, the optimum TCO quickly
energy gain when raising the cooling air temperature. In fact, the rise to 24 °C as it is better to perform WSE iteration with a high
WSE Hourly Average Operation Cost (HAOC) that defines the mean HAOCWSE than IWSE operation.
energy absorption of the cooling plant for one hour under free- From these results it seems interesting to regulate the temper-
cooling operation strongly increase with the raise of TCO. When ature TCO over the year in function of the ambient climate.
TCO = 16 °C. HAOCWSE = 3.55 kW and for TCO = 24 °C.,
HAOCWSE = 4.77 kW. This raise is mostly due to CRAC fan energy 9. The Temperature Adaptive Control Strategy (TACS)
consumption as its airflow increases to maintain the servers under
allowable temperatures. However, concerning IWSE operations, it In this chapter, a new Temperature Adaptive Control Strategy
appears that HAOCIWSE nearly stands still no matter the blowing (TACS) is proposed. The aim of this regulation, is to vary the blown

Table 5
TACS mode selection process.

TACS System modes


TACS free-cooling (WSE) Partial free-cooling (IWSE) Normal
TchHX 6 Tloopmax Tchhot > Tsump AND NOT (WSE) Tsump > Tchhot
 The heat pump is turned off, The heat load  The heat pump is turned on  The heat pump is turned on
is evacuated through the cooling tower  The cooling tower works at 70% of its maximal  The cooling tower works at 70% of its maximal capacity
 The cooling tower works at 70% of its capacity  The 2 by-pass valves on the evaporator V2 and on the
maximal capacity  The 2 by-pass valves on the evaporator V2 and on condenser side V3 are open. The heat exchanger is short
 The by-pass valve on the condenser side the condenser side V3 are closed. The water of the circuited to reduce the pressure loss in the circuits
V3 is closed, the water fully flows through 2 circuits flow through the heat exchanger.  The loop by-pass valve V1 is open, the water fully flows
the IWSE heat exchanger  The loop by-pass valve V1 is regulated, mixing through the HP and the temperature Tloop is set to Tloopmin.
 While TchHX P Tloopmin, the by-pass valve chilled water from the Heat Pump and returning  The blown air temperature set-point is set such as
V2 is open, and the whole water flowrate water to maintain Tloop at Tloopmax Tco = Tloop + 7 °C
goes through the IWSE heat exchanger  The blown air temperature set-point is set such as
 If TchHX < Tloopmin then the valve V2 is regu- Tco = Tloop + 7 °C
lated to maintain Tloop equal to Tloopmin.
 The blown air temperature set-point is set
such as Tco = Tloop + 7 °C
 The loop by-pass valve V1 is closed
B. Durand-Estebe et al. / Applied Energy 134 (2014) 45–56 55

Fig. 11. Weather analysis and TCO evolution with TACS regulation.

Fig. 12. Yearly energy consumption comparison for various TCO set-point and TACS regulation.

air temperature TCO during the year, to perform as much WSE oper- previously said, the cooling tower efficiency varies following the
ations as possible, while getting the lowest HAOCWSE. To calibrate weather conditions. In winter time it is able to maintain low tem-
the system, the boundaries of the blown air temperature (TCO) peratures in the sump that evolve similarly to the external temper-
are set. They correspond to the limits of the previous temperature ature. This tank of cool water allows the system to perform WSE
range with TCOmin = 16 °C and TCOmax = 24 °C. Thus, as a tempera- operations at low temperature and the average TCO set-point
ture drop of 7 °C is maintained between TCO and the loop-circuit remains below 20 °C. When external temperatures goes up, the
water temperature Tloop, the boundaries of Tloop are Tloopmin = 9 °C TCO consign follows the trend to perform WSE operations as much
and Tloopmax = 17 °C. Given these four constants, the system as possible and with the lowest possible temperature. In summer
behaves according to the algorithm presented in Table 5. time the TCO set point is bounded by TCOmax = 24 °C as the CRAC
As it can be seen on Table 5, The TACS system is very similar to fan cannot ensure server safety past that temperature.
the previously described IWSE control strategy. The main differ- Fig. 12 compares the energy consumption of the previously pre-
ence lies in the behaviour of the system during total free-cooling sented simulation with the energy needs required by the system
operations (WSE). When using the TACS regulation, the value of working with the TACS.
the blown air temperature TCO is set in function of the water tem- During the months of January, February, March, November,
perature exiting the IWSE heat exchanger on the evaporator side. December, the installation working with the TACS system perform
Thus during the winter, or in cold periods, TCO is low and the nearly 100% free cooling operation, like the simulation with
HAOCWSE is at its minimum. When the outdoor temperature rises, TCO = 24 °C. However the energy consumption during this period
the cooling tower sump temperature increases as well, leading to is much lower with the TACS system. This is due to the HAOCWSE
a rise in TchHX. In this condition, TCO rises until it reaches the limit that is equal to 4.77 kW for a fixed TCO = 24 °C whereas mean
TCOmax. After this point, if the temperature TchHX continues to rise, HAOCWSE = 3.95 kW for the installation regulated by the TACS with
the system switch to IWSE mode (Table 5 column 2), and the heat a mean TCO = 18.17 °C. The energy gain for these months with a sys-
pump maintain Tloop at Tloopmax and TCO at TCOmax. Finally extremely tem regulated by TACS compared to a single TCO = 24 °C set-point is
hot temperatures might happen during the summer, and the cool- around 18%. When the temperature is rising, the energy gain
ing tower sump temperature might exceed the water temperature decreases and is nearly null during the summer. Indeed, as previ-
exiting the CRAH unit Tchhot. In this case the system switches to ously explained, the system sets the TCO set point as high as possi-
Normal mode with Tloop = Tloomib and TCO = TCOmin in order to mini- ble (24 °C) to maximize the number of hour of IWSE operations.
mise the CRAH fan energy consumption. Fig. 11 plots several infor- From these results, it can be noticed that the TACS regulation
mation regarding an average day and for each month of the year. allowed energy savings during the cold season, but is not efficient
The line with cross markers plots the external temperature and during the summer. Once again this is due to the behaviour of the
the surface around the line is bounded by the maximum and the cooling tower. During the hot season, it is not able to remove
minimum temperature reached during the month. The upper hor- enough energy from the system. Moreover, the TCO set point is
izontal dashed line plots TCOmax and the lower one TCOmin. The con- already at its maximum. Thus the system can only perform IWSE
tinuous line plots the TCO set points that are sent to the CRAC. As operations at the highest allowed temperature. However the
56 B. Durand-Estebe et al. / Applied Energy 134 (2014) 45–56

energy gain over a year compared to a constant TCO = 24 °C set [2] Almoli A, Thompson A, Kapur N, Summers J, Thompson H, Hannah G.
Computational fluid dynamic investigation of liquid rack cooling in data
point is around 7%. Compared to the base case TCO = 16 °C the gain
centres. Appl Energy 2012;89(1):150–5.
is 17%. [3] Patterson MK, Pratt A, Kumar P. From UPS to silicon an end-to-end evaluation
of data center efficiency. In: Digital enterprise group architecture and
planning; 2006 <http://www.energystar.gov/ia/products/downloads/
MPatterson_APratt_Case_Study.pdf>.
10. Conclusions [4] ASHRAE TC9.9 2011. Thermal guidelines for data processing environments-
expended data center classes and usage guidance. In: White paper prepared by
This paper investigates a method to lower the cooling plant ASHRAE Technical Committee (TC) 9.9 mission critical facilities, technology
spaces, and electronic equipment; 2011.
energy consumption by modifying its regulation while keeping [5] Patterson MK. The Effect of data center temperature on energy efficiency. In:
all the IT equipments in the ASHRAE allowable environment. Intel Corporation IEEE; 2008, 978-1-4244-1701-8.
A virtual 32 kW datacenter has been considered to perform the [6] Moss D, Bean JH. Energy impact of increased server inlet temperature APC.
White Paper; 2011, 138.
tests. The infrastructure is cooled by the Integrated WaterSide
[7] Biswas S, Tiwari M, Sherwood T, Theogarajan L, Ching F. Fighting fire with fire:
Economizer proposed by Myer and Sorell [10]. modelling the datacenter-scale effect of targeted superlattice thermal
The air flow and the heat are predicted using the CFD code Thé- management. In: Proceeding ISCA 2011 38th international symposium on
computer architecture. June 4–8; 2011.
tis, while the cooling plant is modelled using TRNSYS. A POD
[8] Durand-Estebe B et al. Data center optimization using PID regulation in CFD
reduced order model using Snapshot technique is created with simulations. Energy Build 2013.
the software MATLAB. It is used in the TRNSYS simulation to per- [9] Demetriou D. Thermally aware, energy-based techniques for improving data
form interpolations between computed CFD results. center energy efficiency. In: Mechanical and aerospace engineering –
dissertations. paper 65; 2012.
Two main effects have been identified. First the increase in [10] Myer A, Sorell V. Integrated waterside economizer operation dor data center
blown air temperature allows the system to raise the evaporator central chiller plants. In: Whitepaper #14 Syska Hennessy Group; 2009.
loop temperature. As a results, the number of WSE or IWSE opera- [11] Bulut H, Aktacir MA. Determination of free cooling potential: a case study for
_
Istanbul, Turkey. Appl Energy 2011;88(3):680–9.
tions increase and reduce the load on the heat pump. Doing so, the [12] Dai J, Das D, Pecht M. Prognostics-based risk mitigation for telecom equipment
HP reduces its energy needs by 78% when the blow air temperature under free air cooling conditions. Appl Energy 2012;99:423–9.
TCO rises from 16 °C to 24 °C. On the other hand, an increase in the [13] Seem JE, House JM. Development and evaluation of optimization-based air
economizer strategies. Appl Energy 2010;87(3):910–24.
CRAC fan flow rate, required to maintain all the servers under the [14] Monfet D, Zmeureanu R. Simulation of a large central cooling and heating
maximum allowable temperature 32 °C is observed and raised the plant using TRNSYS and calibration with monitored data. In: Proceeding 11
fan energy consumption by 43%. Thus the simulations determined international IBPSA conferences, July 27–30; 2009.
[15] Wilo. « catalogue Wilo Technique du bâtiment Chauffage, climatisation
that the optimum cooled air temperature for this system is 24 °C.
réfrigération » <http://www.wilo.be>.
In this configuration, the cooling plant reduces its energy con- [16] Alfa Laval. Thor-LSV Datacenter air cooler. Technical sheet <http://
sumption by 10%, compared to case where TCO = 16 °C. www.alfalaval.com>.
[17] Trane. Fan performance data for central-station air handler. In: CLCH-PRC008-
However, the indicator HAOC shows that operating in WSE
EN; October 2004.
mode at TCO = 16 °C required 26% less energy than at TCO = 24 °C. [18] Mackay K. Energy efficient cooling solution for data centres. In: University of
The simulation with the proposed TACS regulation showed an strathclyde Msc energy system and the environment; 2007.
energy gain of 18% in winter time period compared to a fixed [19] Durand-Estebe B, Lebot C, Mancos JN, Arcquis E. Validation of turbulent
natural convection in a square cavity for application of CFD modeling to heat
TCO = 24 °C set-point. Over the year the cooling plant regulated by transfer and fluid flow in a data center »ASME 2012 ESDA2012-82418.
the TACS system provides an energy gain of 7% compared to fixed [20] Lubin P. Simulation des grandes échelles du déferlement plongeant des
TCO = 24 °C and 17% compared to fixed TCO = 16 °C. vagues. PHD, université de Bordeaux 1; 2004 <http://thetis.enscbp.fr/sites/
default/files/theses/these_lubin.pdf (english)>.
Finally, integration of TACS system does not require any addi- [21] Osmar L. Modelisation du refroidissement des pistons hautes performance.
tional component in the cooling plant. There is only a need to mod- PHD, université de bordeaux; 2012.
ify the main regulation which involves a very low maintenance [22] Goda K. A multistep technique with implicit difference schemes for calculating
two or three-dimensional cavity flows. J Comput Phys 1979;30:76–95.
cost. In returns it can provide up to 17% of electric energy savings [23] Van Der Vorst HA. A fast and smoothly converging variant of Bi-CG for the
on the cooling plant requirement. Hence, it strongly reduces the solution of non-symmetric linear systems. SIAM J Sci Stat Comput
datacenter annual cost, making this system economically 1992;44:631–44.
[24] Nicolas X, Medale M, Glockner S, Gounand S. Benchmark solution for a
competitive.
threedimensional mixed convection flow – Part 1: reference solutions. Num
In the future, further simulations will be carried out with blade Heat Trans 2011;60(5).
server introduced in [8]. It will probably require different control [25] Nicolas X, Medale M, Glockner S, Gounand S. Benchmark solution for a three-
dimensional mixed convection flow – Part 2: analysis of Richardson
strategy and optimal TCO set-points. Furthermore the effect of Air
extrapolation in the presence of a singularity. Num Heat Trans 2011;60(5).
Distribution System modification on the cooling plant consump- [26] Bui-Thanh T, Damodaran M. Proper orthogonal decomposition extensions for
tion will be investigated. New POD reduced order model will be parametric applications in transonic aerodynamics. In: Proceeding of 15th
created with cold aisle containment and the effect on cooling plant AIAA computational fluid dynamics conference; 2003.
[27] TRNSYS 16 Documentation. Volume 5 mathematical reference
energy consumption will be calculated. <www.trnsys.com>.
[28] Thermal Energy System Specialist (TESS). TESS library documentation
<www.trnsys.com>.
References [29] Aström KJ, Wittenmark B. Computer controlled systems. 2nd ed. Englewood
Cliffs, NJ: Prentice Hall; 1990 [ISBN 0-13-168600-3].
[30] Aström KJ, Hägglund T. PID Controllers: theory, design and tuning. 2nd
[1] Cho J, Lim T, Sean Kim B. Measurement and predictions of the air distribution
ed. International Society for Measurement and Control; 1994 [ISBN 1-55617-
systems in high compute density (internet) data center. Energy Build
516-7].
2009;41:1107–15.

View publication stats

You might also like