As many parameters of the cooling system in data center mainly depend on manual control, it is urgent to fix the problem of simulation and optimization at the same time and develop effective mitigation methods. In this paper, we proposes a combined method based on machine learning. Firstly, we have built the neural network with the temperature and humidity of the equipment in real data, then established evaluation model, and finally the optimal setting has been obtained by a genetic algorithm. Our results show that that the effect is better than artificial method and traditional greedy algorithms with 3% - 15% relative reduction in errors of temperature and humidity per set of parameters.
Data center is a facility that consumes a lot of energy and there is a strict standard of
environmental factors. For example, if the temperature is too high, the life of the
component will be reduced, and at the same time, the discharge capacity of the battery will
be lowered if the temperature is too low. Early physicists tried to solve this problem by
analyzing the structure of the cooling system and combining the formulas of fluid
mechanics [
At present, the methods of controlling temperature and humidity in a data center are
mainly focused on load balancing or changing the topology of equipment and air
conditioning through Computational Fluid Dynamics(CFD) to simulate [
In recent years, deep neural networks have been widely applied in various fields. In
particular, Google's approach to reducing Power Usage Effectiveness(PUE) through
neural networks in 2014 [
After getting an accurate simulation model, we need to get an optimal combination of parameters to make the temperature and humidity meet the standard. It is usually more convenient and accurate to use the machine learning algorithm to calculate the optimal solution. The advantage of the genetic algorithm is that it can deal with constraints very well, and jump out of local optimum easily, and finally get the global optimum solution. That is, the global search ability is strong. For the setting of the fitness function, we set up a hierarchical and segmented evaluation function on the basis of the above neural network model in order to take better care of the values with further distance with standards.
In this paper, we propose a combined machine learning model to overcome the shortcomings of a single algorithm or manual control of cooling parameters in data centers. The paper is mainly divided into the following parts: Section 3.1 describes how to use the neural network to simulate the actual production environment; Section 3.2 explains the method of increasing the weight of Abnormal values by layered function; Section 3.3 explains how the genetic algorithm is applied to the calculation of optimal solution; Section 4 describes some formulas and realizations in the implementation of the algorithm. Finally, the Section 5 will summarize the paper and explain the future work. 2 2.1
To cope with these problems, the neural network is chosen as the framework of a data
center model. The neural network is a machine learning algorithm which imitates the
interaction between neurons in cognition. Like most learning systems, the accuracy of
the model increases with the addition of new training data at each iteration. Google is
a pioneer in intelligent data centers. They have used a fully-connected neural network
to simulate data center energy consumption [
Our decision to optimize temperature and humidity comes from [
Google uses neural networks to get simulated data from data centers, and uses control variables to predict the label trend when one or two variables change simultaneously. Although the method similar to the greedy algorithm can reduce PUE directly, it is easy to fall into local optimal solution and the analytical ability when sets of variables change simultaneously is weak.
In order to solve this problem, the genetic algorithm is adopted in this paper. Genetic
algorithm is based on the evolution process of nature: selection, crossover and mutation,
through the fitness function to evaluate the possibility of individual survival. The
Genetic algorithm is widely used in finite element problems, such as in the green cloud
memory allocation strategy, compared with the traditional SLC configuration, the
maximum memory usage based on genetic algorithm can be reduced by 76.8% [
At the same time, due to the controllability of genetic algorithm, we adopt an
adaptive method: by artificially restricting the survival probability of operators [
In the application of actual scenes, a single model often cannot solve the problem of
complex and multi-constraint conditions [
In the experiment, our dataset contains more than 300 dimension numerical or
nonnumerical features. Referring to the method of [
In addition, it should be noted that the standardization and normalization of data are also very important. In our case, the difference between the order of magnitude of the data items is very large. In this case, the standardization and normalization of data can speed up the convergence rate, reduce the calculation cost and improve the accuracy rate. In this article, we have mainly adopted the following methods:
The first step is to normalize the data to [
x - MEAN(x) xnorm =
MAX (x) - MIN (x) (1)
Secondly, the batch normalization layer is added after the input layer, and the
following formula is used[
å w2 2n w 2
Where µB is the mean and sB is variance. g is a scaling variable and b is a translation variable. In the training process, the variance and mean are calculated by training input data, and g and b are the parameters that need to be learned. This ensures that every time the data is normalized, it still retains the trained features.
At the same time, it can complete the normalization operation and accelerate the training. At each level, we also add the L2 regularization term. The formula is as follows: l (3)
Where c0 is the original output term, w is the product parameter of all neural networks, l is the normal term coefficient, and n is the sample size of the training set. The addition of regular term will reduce the weight of, reduce the complexity of the network and prevent over-fitting.
xˆi ¬ xi - µB sB2 + e yi ¬ γxˆi + β º BNγ,β (xi ) (2) ifxi £ 0
Finally, in order to solve the problem of inadequate expression of the linear model,
we add activation function to make neural network can deal with complex data better.
In the middle hidden layer, we use the PReLU function[
The reason is that it has a faster convergence speed than tanh/sigmoid and does not appear the problem of gradient disappearance. At the same time, when x is less than 0, a smaller gradient (ai) is used to replace 0, which can solve the problem that the excessive gradient stops activation after passing through the ordinary ReLU unit. After the last level of output layer, we adopted tanh:
1 - e-2x tanh(x) = 1 + e2x (5)
Because the output range is [
Based on the above theoretical analysis and experimental results, we select 100 parameters as input, 336 labels (168 temperature and 168 humidity measurement points) as output, including five hidden fully-connected layers, each containing 50 nodes. Our final network structure is as follows: We obtained the standards of temperature and humidity of each monitoring point from long-term experience and expert markers, and found that if use mean-square error(MSE) as the evaluate function of the genetic network directly, it could not be very good to optimize the poor performance measurement points. And in the actual production situation, we pay more attention to the data far from the standard value, so we propose a hierarchical evaluate function to optimize the combination model.
The hierarchies mainly include two meanings, one is the numerical hierarchies of different distances from the standard value, the other is the weight distribution of different monitoring points. Firstly, the algorithm for different distance thresholds: There is a manually calibrated standard threshold set. Each standard of threshold has a pair of upper and lower limits. The internal is the length of the threshold internal in which the value (temperature or humidity) located. The abs_value is the absolute value of distance from value to the inner standard.; basei and weighti are the hierarchical base value and weight value, and the algorithm will eventually output a score. When the value is closer to the standard, the score is bigger.
score = basei + weighti (1 - abs _ value / internal) (6)
Secondly, due to the multi-level and multi-quantity characteristics of the monitoring points, the weights of the measuring points for different scores should also be different: the input of the function is the set of multi-point scores obtained by formula (6); These values are divided into low_score and high_score according to a threshold. then low_pro and high_pro are calculated from normalized sets of the values as the proportion of low and high scores. The method of normalization is l2 normalization. score = low_ score*low_ pro + high _ score*high _ pro (7)
These two functions are very meaningful in actual production, because the temperature or humidity close to the standard value does not need to be processed, and when it exceeds a certain threshold may lead to safety accidents. 3.3
For genetic algorithms, the most important steps are selection, crossover and mutation. The step to produce the optimized offspring is the selection. According to individual fitness and certain rules, the selection algorithm selects some individuals with good traits from the nth generation population and inherits them into the next generation (n + 1) population. In this selection process, the greater the individual's fitness, the greater the chance of being selected to the next generation. The fitness ratio fi of an individual i and NP of population size is the probability formula for i to be selected as follows: Pi =
fi åiN=P1 fi (i = 1,2,3...., NP) (8)
It is obvious that using the fitness function optimized by us can greatly reduce the survival probability of operators with large differences in individual values, and allow all operators with small or moderate universal distances to survive.
In addition, in order to speed up the convergence speed and generate the operator with
higher fitness, we adopt the idea of the adaptive genetic algorithm[
pˆc = p（c1 pˆm = p（m1 f ¢ - favg fmax - fmin f ¢ - favg fmax - fmin ） ）
(9) (10)
Where pc and pm are the standard probability of crossover and mutation, and f ¢ is the fitness of the individual , favg , fmax , fmin are the average, maximum and minimum value of fitness of this batch of population. When f ¢ is higher than favg , pˆm and pˆc will be lower to keep the excellent individual unchanged.
Finally, the whole process of optimization system is shown in Fig.3..
Input Invariable
Parameter
Initializing Variable
Parameters
Neural Network Simulation model
Hierarchical evaluation function Mutation Crossover
Select
Adapt Possibility of Crossover and Mutation
No
Stopping Criteria？
Yes Output Optimized
Parameter An accurate cooling system simulation is very useful for data center management. In order to ensure the safe operation of the equipment, the cooling parameters of the actual production environment are usually not directly adjusted. For example, after a system reboot, because of the great changes in the external environment, administrators can simulate according to the neural network to understand the general range of each parameter, and then adjust, it will not produce security risks, but also save a lot of manpower and material resources.
Another advantage of neural networks is that when the amount of data increases, the simulation results are more accurate than traditional machine learning methods such as decision trees. Because of this characteristic, we boldly select 100 parameters in many parameters, most of which are input from the network. At present, our experimental results are as follows:
Using the hierarchical function does not reduce the overall error compared to ordinary MSE, but our goal is not to adjust the better points, but to pay more attention to the worse points. From Fig.7 (normalized error distribution) we can see that we did reduce the point farthest (normalized MSE>0.05)from the standard value. The remaining points that are not optimized are because the adjusted parameters are not sufficient for this purpose, such as the distance of the measurement points being too far from the topological distance of the air conditioner being adjusted. As mentioned above, administrators can set the cooling parameters according to their own experience and the auxiliary simulation of the algorithm, but the disadvantages of doing so are obvious: time-consuming and laborious, and the accuracy is not high. This shows that a complete set of AI algorithms is necessary. Due to the limitation of time and the actual needs of the experiment, we only change 8 parameters, and fix the other 92 parameters for the experiment. These 8 parameters are the temperature and humidity control of 4 of the air conditioners. In this paper, we choose the artificial method and greedy algorithm to compare with our method. Setting parameters manually according to experience or based on physical models is a common practice in the industry, and Google's use of neural network sensitivity analysis to adjust the parameters of optimization parameters is basically greedy algorithm. In theory, the complete algorithm comparison should be in the simulation phase and the optimal solution phase using different algorithms, such as "neural Network + genetic algorithm" and "CFD + artificial." But because of the amount of calculation, we use the neural network in the simulation phase uniformly, but still can see that our method has a obvious effect. In addition, because only 8 parameters are changed, the difference between the algorithms is sometimes small, This because other parameters besides them need to be adjusted. And with the number of variable parameters increases, the greedy algorithm is easy to fall into the local optimal solution, while the effect of experience factors in the artificial method will be weakened as well, in the actual production environment, our method will be more excellent.
In the experimental details we refer to the method in [
In the experiment, our fitness function uses the hierarchical cost function, so the final result of MSE may not be the best, but in order to show the results more clearly, our final evaluation method is still MSE, we can see that the effect is still better than other methods which is shown in TABLE I (the Time is 5 points randomly selected).
To quantify this effect in the actual production situation, assuming that the temperature average error is 2 ℃when manually adjusting a certain set of parameters, and 10% error optimization is calculated according to the experimental result, then the temperature optimization for all devices is 2*10%*168= 67.2 (℃). Since we use a hierarchical evaluation function, most of this optimization come from poorly performing points which has significant benefits for safe operation of equipment. Moreover, this is calculated without taking into account the humidity error.
Time 1
Artificial Method 2.2894
In this paper, we propose a combined algorithm based on machine learning to optimize
the cooling system of data center. This approach can not only save human resources
and time, but also make the control of cooling parameters more accurate. At the same
time, it avoids the defect that the traditional greedy algorithm is easy to fall into the
local optimal solution, and provides a better solution for multi-parameter simultaneous
adjustment. However, due to the limitation of data set, we failed to further explore the
relationship between setup and energy consumption. Besides, the lack of
interpretability of neural networks is also a difficult problem. For this paper, the most important
problem of the lack of interpretability is that it is impossible to explore the effect of
adjusting cooling parameters on the measurement points at different topological
locations. We have learned that in the work of [
This work is supported by Science and Technology Program of State Grid Corporation of China under Grant No.5442DZ170019.