WEB SERVER RELIABILITY PREDICTION Răzvan-Daniel ALBU ...

34 downloads 118 Views 425KB Size Report
KEYWORDS: prediction, reliability, recurrent neural networks, web server. 1. ... The Dotcom-Monitor software is ... addition, Dotcom-Monitor provides a large.
Nonconventional Technologies Review – no. 1/2011

WEB SERVER RELIABILITY PREDICTION Răzvan-Daniel ALBU Oradea University, Faculty of Electrical Engineering and Information Technology, Universităţii Street 1, Oradea, Romania, [email protected]

ABSTRACT: The application described in this paper consists of monitoring the activity for the web servers in order to collect the specific data. We accumulate different data sets from monitoring the web server’s activity and for each one we predict the server’s reliability using the proposed RNN architecture. Predicting an error with severe consequences for the performance of a server requires measuring the capacity of a server at any given time. This measurement is highly complex, if not impossible.

KEYWORDS: prediction, reliability, recurrent neural networks, web server.

1. INTRODUCTION

have as an effect people waiting for too long and/or even leaving the website. When a server performs slowly, it does not work at all, so no records will be created in error logs [4]. Thus, without monitoring the server’s performance, you will never know that people experience problems when they access your site [6, 7, 9]. The Dotcom-Monitor software is able to check a number of parameters. In addition, Dotcom-Monitor provides a large number of online, real-time and e-mail reports and charts. The reports we have considered as most interesting for our problem are presented below: • Device Response Time Chart: This chart provides aggregated summary for all website monitoring locations on device performance. • Distribution by Day of the Week Chart: This monitoring chart aggregates statistics performances by day of the week and shows it in the Bar chart. • Distribution by Hour of the Day Chart: It is a similar chart to "Distribution by Day of the Week", only the data is aggregated hourly. The chart shows the website performance over 24 hours. • Monthly Summary E-mail Reports: Once again, similar to weekly and daily reports, only that it incorporates statistics for a whole month. Those reports are e-mailed the first day of the month including the data for the previous month.

In this paper we propose a neural network prediction model (NNPM), based on the data collected from an experimental Web Server system. The results demonstrate the ability of the recurrent neural networks (RNN) for obtaining reasonable results. Errors are discovered before they become critical and interrupt the normal activity of the web server. This will certainly improve the system’s reliability. The method is based on the correlations observed between the variables measured during the server’s activity and on the observation that a piece of software passes through multiple deterioration processes before it fails [1, 3, 5, 11]. If these deterioration processes can be monitored and predicted, optimal maintenance actions can be scheduled in order to improve the reliability of the system. Neural network (NN) methods have been considered to be a very promising approach to the reliability prediction, due to the agility of the nonlinear and arbitrary function’s approximation. The NN methods take into consideration the deterioration processes instead of assuming the mathematical model for the damage propagation. 2. RELIABILITY OF WEB SERVERS Web server reliability depends on a number of factors: web server configuration, web server CPU load, the number of services running on the server and the number of users accessing the server at a given moment. Any of these factors may cause poor web server performance, which will

3. NEURAL NETWORKS 3.1 Why RNNs? Neural networks have been successfully applied for a long time in various domains

5

Nonconventional Technologies Review – no. 1/2011

such as research, business and industrial environments. They have been used to solve various problems like: data prediction, classification and function approximation. Many of the real-world problems which one might think would require recurrent architectures have solvable solutions by using layered architectures. For this reason, we would advise engineers to try, first, the layered architecture before resorting to the ―big gun" of recurrence. The recurrent networks are often avoided because of the fear for unreasonable learning hours and incomprehensible algorithms and mathematics. So, there is no reason to use a recurrent network when a layered architecture suffices. On the other hand, if recurrence is needed, there is the availability of an overabundance of learning algorithms. The reason for exploring recurrent architecture lies in their potential for dealing with temporal behaviours. Although, there still rises the question: ―Why RNNs for web server reliability prediction when we already have so much layered architecture that could solve this problem?‖ The relative superiority of recurrent networks to feed-forward networks in forecasting is not just due to its ability to model time series data with lower errors, but rather to model a parsimonious training set [2]. With the rapid growth of processing speed in the last years, the context in which we define an efficient method changed. The capabilities of many digital electronic devices are strongly linked to Moore's law. Processing speed, memory capacity, sensors and even the number and size of pixels in digital cameras are improving at exponential rates. This exponential improvement has given us a new perspective on the prediction problem. Nowadays the learning time, a parameter that was critical in designing an efficient NN based solution, is not such a big problem like 10 years ago, because we already have on the market processors similar to Intel Core i7 990x that can reach speeds of 4.5-5 GHz. Consequently, we propose a model for the problem of reliability prediction focused more on the prediction accuracy then the learning time. Zhigang Tian and Ming J. Zuo recommend a neural network prediction model called extended recurrent neural network (ERNN). They developed an ERNN based system for health state prediction of

gearboxes. This is based on the vibration data collected from an experimental system. The results demonstrate the capability that the ERNN has for obtaining satisfactory health state prediction results. Charles Thompson in the paper: ―Fighting Virus and Hack Attacks with a Network Analyzer‖ shows how a network analyzer, historically used for network troubleshooting, can also be used for defending against the security threats. Fulya Altiparmak, Berna Dengiz and Alice E. Smith, in the paper ―A General Neural Network Model for Estimating Telecommunications Network Reliability‖, put forth a new encoding method that uses neural network models for estimating the reliability of telecommunication networks with identical link reliabilities. 3.2 Learning with fixed points One problem with fixed points is that recurrent networks do not always converge to them [8, 10]. However, there are a number of special cases that guarantee convergence to a fixed point. Some simple linear conditions on the weights such as zero-diagonal symmetry (wij=wji, wii=0) guarantee that Lyapunov function decreases until a fixed point is reached:

Wij are the weights of the connections from unit i to unit j and yi is the activation level of unit i:

And the general equation that model the neural network is:

Where Ǿ is an arbitrary differentiable function and Ii are the inputs. Aliya has showed that a unique fixed point is reached, regardless the initial conditions, if:

6

Nonconventional Technologies Review – no. 1/2011

networks. The Elman context layer differs from the original Elman RNN because the two context neurons obtain inputs from the output of the hidden layer after a delay of one time unit, and from itself. In the Jordan context layer the difference is that the context neurons obtain inputs from the output error of the network after a delay of one time unit, and from itself. In both context layers there are two neurons with self-feedbacks. For predicting time series in the output layer, we need just one neuron. We use also two neurons in the input layer because it has been reported that every data point in a time series is only strongly dependent on the immediate past two values [12, 13, 14]. The linear activation function is used in the output layer, the Jordan context layer and the Elman context layer. The sigmoid activation function is used in the hidden layer.

Other empirical studies show that applying fixed point learning algorithms stabilizes the networks. However, the fixed point learning algorithms can still have problems even when it is guaranteed that a network settles to a fixed point. The learning procedures compute the derivative of some error measure. This gradient is used by an optimization procedure in order to minimize the errors. The optimization procedures assume that the mapping from the network's internal parameters to the resulting errors is continuous and can fail when this assumption is violated. This means that the learning algorithm changes the locations of the fixed points by varying the weights. Therefore, it is also possible for a result to stumble upon such a discontinuity. This will induce errors which will appear suddenly. 4. THE PROPOSED RNN

4. CASE STUDY In support of time series prediction, a neural network can be considered for a nonlinear mapping between a subset of the past time and future time series values. Studies confirm that NN approaches perform better than statistical autoregressive moving average (ARMA) methods. In addition, recurrent neural networks (RNN) perform even better for the time series predictions [15, 16]. The proposed architecture is presented in Fig. 1.

4.1 Neuro Solutions 6 Neuro Solutions is a graphical neural network development tool, which can easily create a neural network model for the input data. This software combines a modular design interface with advanced learning procedures. Neuro Solutions provides the power and flexibility needed to design the best neural network for our problem.

Figure 2 Neuro Solution Breadboard 4.2 Prediction Results Our experimental system consists of following items: • The web server: IBM x3300 system with 2.2 GHz processor, 2 GB DDRAM, 160GB SATA hard disk, Fast Ethernet LAN adaptor.

Figure 1 The proposed RNN The RNN has two context layers: the Elman context layer and the Jordan context layer, both with some differences from the original Elman and Jordan recurrent neural

7

Nonconventional Technologies Review – no. 1/2011

• The Neuro Solutions 6 application for simulating the proposed RNN. • The Dotcom-Monitor applications for collecting training data. Based on Dotcom-Monitor reports we can construct our data set which will be used as training data set for the RNN. For example, if we use the Monthly Summary Email Reports presented in Fig.3, as training data set for our RNN, with a time-window of 5 days, the RNN will predict an availability of 99.43% on 10/11/2009. As we can see the prediction result is very close to the monitored value that shows an availability of 99.82%. This indicates a network prediction performance rate of 99, 6%. This is our network’s best prediction score. The 5 days time window means that we build a RNN with 5 input neurons. Each one of them has as input the availability monitored on the respective day. We train the RNN inputting values for 5 days starting with 9/11/2009. The second input data set represents values for 5 days, starting with 9/12/2009, and so on. We expect as output for each training data set, the monitored value on the 6th day. A monthly summary email report is presented below (Fig 3.).

Figure 4. Distribution by Hour of the Day Chart

Figure 5. Distribution by Day of the Week Chart

Figure 6. Device Response Time Chart In a similar way, we use data from other reports to create different training sets and thus more predictions. Using the Distribution by Hour of the Day Chart we can construct data sets for predicting days when we have more visitors and more load. The RNN prediction performance is illustrated in Table 1. For each Dotcom-Monitor report we test our network with three different data sets and measure the prediction percentage. As we can see, the best prediction results are obtained using data from the Monthly Summary Email report where the RNN shows an average performance of 98.86%. The lowest prediction score is given by the Distribution by Hour of the Day report, where our RNN shows only an average performance of 87.3%. The results demonstrate the capability of the RNN for producing satisfactory prediction results. A high-quality health state prediction system can help the administrator to learn about the possible errors or downtimes in advance; this could actually improve the reliability of a Web Server.

Figure 3. A Monthly Summary Email Report

8

Nonconventional Technologies Review – no. 1/2011

Table 1 The RNN prediction performance Report Device Respo nse Time Chart Distribution by Day of the Week Chart Distribution by Hour of the Day Chart Monthly Summary Email Reports

Epochs/Experiment, the Exemplars/Epoch and the Samples/Exemplar. In this particular example we will select as parameters 1000, 5 (since you have a 5 days’ time-window) and 10. This means that each sample will be repeatedly presented to the network 10 times, to let the output relax. This parameter is crucial for learning. If the network is not relaxed enough, the output activation will not be in the steady state value and will produce an erroneous error estimate. In our opinion, one of the reasons why recurrent systems are not as popular as static systems is the difficulty of dynamic learning.

The RNN average prediction performance DataSet1 DataSet2 DataSet3 Avg% 86.2 88.1 89.5 87.9 % % % %

91.3 %

98.3 %

98.3 %

95.9 %

87.6 %

85.3 %

89.1 %

87.3 %

98.3 %

99.6 %

98.7 %

98.8 %

5. CONCLUSIONS In this paper, we propose a RNN prediction model. This model is developed for the health state prediction of a web server based on the data collected with DotcomMonitor software. The results demonstrate the capacity of the RNN to produce at least satisfactory prediction results. The RNN we recommend, combined with the DotcomMonitor software, could be very good tools for every network administrator. A good health state prediction can help the network administrator be aware of possible errors or downtimes in advance. This could really improve the reliability of the web server.

In order to implant fixed points into recurrent systems, the backpropagation technique is used. The fixed-point learning is considered in addition to the static backpropagation. In fixed-point learning, the first action is the forward propagation of the activations. This procedure repeated for a certain number of times will induce the relaxation period. This has to be repeated until the network reaches its own dynamic. After the network stabilization an error can be computed at the output. Then, the error is propagated backwards through the network. Once again this must be done several times in a row, in order to allow the network to relax. After relaxation, the error at each output can be multiplied by the relaxed activation for updating the weights. We have to select the relaxation time both in the forward and backpropagation phases. Using Neuro Solutions 6 we can select the Fixed Point radio button in the Dynamic Control Inspector (Fig.7).

ACKNOWLEDGEMENTS This work was partially supported by the strategic grant POSDRU/CPP107/DMI1.5/S/80272, Project ID80272 (2010), co-financed by the European Social Fund-Investing in People, within the Sectorial Operational Programme Human Resources Development 2007-2013.

6. REFERENCES [1] AHMADI, F., & VALADAN, Z. M. J, The Application Of Neural Networks, Image Processing And CAD-Based Environments Facilities In Automatic Road Extraction And Vectorization From High Resolution Satellite Images, The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Beijing, pp. 37, (2008). [2] ALBEANU, G., MADSEN, H., POPENTIUVLADICESCU, FL., DUMITRU MARIUS, An extendable Java-based tool for time series analysis, ENBIS 10 Antwerp, 12-16 September 2010. [3] ALTIPARMAK F., DENGIZ B., & SMITH, A.E., A General Neural Network Model for

Figure 7. The Dynamic Control Inspector Within the Dynamic Control Inspector there are 3 forms of cells: the

9

Nonconventional Technologies Review – no. 1/2011

[11] GEBRAEEL N., & PAN, J., Prognostic degradation models for computing and updating residual life distributions in a time-varying environment, IEEE Trans. Reliability, vol. 57, no. 4, pp. 539–550, (2009). [12] GEBRAEEL, N., ELWANY A., & PAN, J., Residual life predictions in the absence of prior degradation knowledge, IEEE Trans. Reliability, vol. 58, no. 1, pp. 106–117, (2009). [13] GROTTKE M. & TRIVEDI, K. S., Software errors, software aging and software crejuvenation, Journal of the Reliability Engineering Association of Japan, vol. 27, no. 7, pp. 425–438, (2005). [14] GROTTKE, M., LIE, L., VAIDYANATHAN K., & TRIVEDI, K., (2006), Analysis of software aging in a web server, IEEE Trans. Reliability, vol. 55, no. 3, pp. 411–420, September 2006. [15] HALFOND, W. G. J., ORSO A., & MANOLIOS, P., WASP: Protecting Web Applications Using Positive Tainting and SyntaxAware Evaluation, Software Engineering, IEEE Transactions. vol. 34, no. 1, pp. 65–81, (2008). [16] HOFFMANN, G., & MALEK, M., Call availability prediction in a telecommunication system: A data driven empirical approach, presented at the IEEE Symposium on Reliable Distributed Systems (SRDS 2006).

Estimating Telecommunications Network Reliability, IEEE transactions on reliability, VOL. 58, NO. 1, (2009). [4] ARLITT M., & WILLIAMSON, C., Understanding web server configuration issues, Software—Practice and Experience, vol. 34, no. 2, pp. 163–186, (2004). [5] BAO, Y., SUN, X., & TRIVEDI, K. S., A workload-based analysis of software aging and rejuvenation, IEEE Trans. Reliability, vol. 54, no. 3, pp. 541–548, (2005). [6] BAYUK, J. L., Productive intrusion detection, Comput. Security J., vol. 18, no. 3–4, pp. 23–32, Summer–Autumn, (2002). [7] BECHER, M., Web Application Firewalls, Applied Web applications security. Berlin, (2007). [8] BOLCH, G., GREINER, S.H., & TRIVEDI, K.S., Queueing Networks and Markov Chains: Modeling and Performance Evaluation with Computer Science Applications (The White Book), second ed. New York, NY: John Wiley, (2006). [9] CHARLES THOMPSON, Fighting Virus and Hack Attacks with a Network Analyzer, IEEE transactions on reliability, VOL. 59, NO. 1, (2010). [10] EDUARDO O.C., RAMIREZ POZO, A.T., & VERGILIO, S.R., A Genetic Programming Approach for Software Reliability Modeling, IEEE transactions on reliability, VOL. 59, NO1, (2010).

10

Suggest Documents