Thermal Management of a Data Center White Space

Abstract

The total
electrical energy consumption by all the operational data centers (located all
over the world) is enormous (approx. 1% of the global electricity demand
20000TWh). This electrical energy is required 24x7 to operate and cool all the
IT equipment present in the data center. The electrical energy required to cool
all the servers present in the white space can range from as low as 10% to as
high as 40% of the total data center electrical consumption. The server's inlet
temperature has to be within the ASHRAE recommended range (18o C – 27o C), so
that they can function correctly.   A
simplified design of a raised floor white space with hot aisle / cold aisle
configuration is considered. The tile flow rate through the floor tiles
influences the server's inlet temperature. To control the tile flow rate, 11
design variables of the data center white space are identified. These are the
position of the 4 perforated plates, the amount of perforations of each
perforated plates, the floor tiles perforation, the raised floor height, and
the CRAH distance to the cabinet. 600 design samples of the white space are
generated by applying the Latin Hypercube Sampling (LHS) technique on these 11
design variables. A well-validated CFD software 6SigmaRoom by Future Facilities
is used to generate the CFD results. The standard k-ε model is used to model the turbulence in the
CFD simulations, in a steady-state condition. A database of 600 samples is
generated by recording the cabinet's inlet temperature, flow rate of floor
tiles from the CFD simulations, and the corresponding changeable design
parameters generated by the LHS technique. The error due to CFD simulation is
estimated at less than 4% for the tile flow rate and 1.7o C for the server
inlet temperature.  Four Artificial
Neural Networks (ANN) are trained on the data from the database to predict the
floor tile's tile flow rate and the cabinet's inlet temperature, respectively.
The average R2 prediction (testing) accuracy is 0.97 for the tile flow rate
predictions and 0.92 for the cabinet's inlet temperature predictions. Their
average prediction error is less than 5% for the tile flow rate and less than
20 C for the cabinet's inlet temperature. The Non-dominated Sorting Genetic
Algorithm-II (NSGA-II), a variant of the genetic algorithm, is used to find the
optimum values of the 11 design parameters of the white space. These optimum
values are going to ensure the server's inlet temperature to be within the
ASHRAE recommended range. The genetic algorithm optimizes the design variables
based on the predictions made by the neural network predicting the tile flow
rate. The values of the optimized design parameters are verified using the
6SigmaRoom software by comparing the server's (mean) inlet temperature of the
optimized case with the non-optimized case. More number of servers have their
(mean) inlet temperature below 270 C in the optimized case as compared to the
non-optimized case. The electrical power required by the CRAHs to cool the
white space is reduced by 10% in the optimized case as compared to the power
which is required by the CRAHs in the non-optimized case. In this study, a CFD
simulation of the white space took 40 minutes. The neural networks took less
than a minute to make the predictions, and the NSGA-II algorithm took less than
10 minutes to find the optimized design parameters of the white space.  Thus, in this thesis, it is shown that using
an artificial neural network and a genetic algorithm, in combination with
computational fluid dynamics gives satisfying results in optimizing the white
space design, required to keep the server's inlet temperature within the ASHRAE
recommended range. The computational time required to find the optimum white
space design is also reduced by using a neural network and a genetic algorithm.
The prediction by the neural network and the optimization performed by the
genetic algorithm can be improved further with the availability of more
training data and in-depth knowledge of applying these techniques (the neural
network and the genetic algorithm) in predicting and optimizing the solutions
respectively.