In this research, we propose, sign, and implement the NAP (Network Bandwidth Predictor) for rapid network performance prediction. NAP is a new system that employs a neural network based approach for network bandwidth forecasting. This system is designed to integrate with most advanced technologies. It employs the NEWS (Network Weather Service) monitoring subsystem to measure the network traffic, and provides an improved, more accurate performance prediction than that of NEWS, especially with applications with a network usage pattern.
The NAP system has been tested on real time data collected by NEWS monitoring subsystem and on race files. Experimental results confirm that NAP has an improved prediction. Index Terms? Performance prediction, Network bandwidth, Artificial Neural Network, Distributed computing 1. Introduction Performance monitoring and forecasting is an active area of research. In the growing world of networking, more emphasis is being placed on speed, connectivity, and reliability.
When network problems occur, they often result in catastrophic breakdowns. Due to the heterogeneity and the constantly varying nature of the network traffic, there are only a few works available to revive prediction of network performance in terms of available bandwidth and latency in a heterogeneous environment, such as Grid computing. Network Weather Service (NEWS) [1, 2] is a welled network performance measurement and prediction system in Grid computing.
However, its simple prediction methods cannot satisfactorily capture the complicated short and long-range temporal dependence characteristic of heterogeneous network traffic. In this paper, we design and implement NAP for online performance forecasting that uses artificial neural network based mechanism for bandwidth redaction. NAP forecasts the available bandwidth, the maximum rate that the path can provide to a flow, without reducing the rate of rest of the traffic in the path.
The system employs NEWS for traffic monitoring and a neural network based approach for bandwidth predictions. NAP is more powerful than simple statistical models and exhibit excellent learning ability. The system can capture complex communication patterns more effectively. NAP employs a non-linear representation that alleviates the problems of linear models. Section 2 discusses the overall architecture of the NAP implementation; Section 3 presents the experimental results with the data collected using the NEWS monitoring system.
Section 4 concludes and summarizes the work and discusses future research. 2. Component Design and implementation of NAP We propose the design of the Network Bandwidth Predictor (NAP) system for network bandwidth prediction. NAP has the following different interacting modules: Network Traffic Monitor, Network Traffic Pre-processor, Network Traffic Predictor, Database and a GUI that work together to provide network traffic forecast. NAP uses NEWS for network traffic monitoring and uses neural network for forecasting.
Network traffic data are compiled into a database to serve as inputs to the neural network based forecasting model. Each sensor periodically takes a performance measurement from the network link it is monitoring and stores the bandwidth information with a time stamp in the database. The resulting collection of measurements, ordered by time stamp, forms a training input to the forecasting subsystem (Neural network based) which generates a prediction of what the performance will be during a given time period. 2. 1.
Network Traffic Monitor The network traffic must be periodically monitored and collected. NEWS monitoring sub-system provides this functionality. It is an active measurement methodology that estimates the hop-by-hop available bandwidth between a source and the destination node on a single link. To monitor network links, the NEWS conducts end-to-end network probes. One host opens a connection with another and sends a small message to measure the link round-trip time, and sends a large message to measure the throughput.
Setting up the NEWS to start monitoring the network traffic involves starting the NEWS daemon processes that provide rectory services, persistent storage, resource monitoring, and forecasting. Each of these hosts listens for service requests on a particular port. NEWS host can be specified by giving the underlying machine and the port it’s listening to. For measuring the network’s TCP traffic, the TCP message monitor is the skill that needs to be employed. This skill monitors the TCP bandwidth and latency beјen each pair of a set of machines . 2. 2.
Network Traffic Pre-Processor The main functionality of the network traffic data pre-processor is to process the raw traffic data elected by the NEWS subsystem and produce a processed training data set This component has been implemented in java. The protestant API takes the input file that contains the raw traffic data collected by the network traffic monitor and the bin size which is the frequency at which the user wants to make predictions as input and output a file that contains the timestamp, minimum, maximum, and average number of bits in one second in the specified bin size.
The output is then passed into the Congregationalists API which then produces the final input that contains al the information (timestamp, month, day, hour, minute, minimum number of bits in one second in that bin size, maximum number of bits in one second in that bin size, average number of bits in the past 1 row, 2 rows, 6 rows, average number of bits in one second in that bin size and the value to be predicted in step one) that are needed for neural network training. 2. 3.
Network Traffic predictor The main functionality of this component is to give a forecast of the network traffic using the traffic history collected by the Network Traffic Monitor. We use neural networks, with their remarkable ability to earn from examples and derive meaning from complicated or imprecise data, to extract patterns and detect trends of available bandwidth. Thus, this component uses multilayer perception neural network with backstopping training in order to make the predictions.
This component of NAP uses weak, a collection of machine learning algorithms for solving real-world data mining problems This component has been implemented in Java. This component creates a neural network. The neural network is then trained  using the processed data obtained by the network traffic pre-processor. The trained neural network is then used to make the actual predictions. The BandwidthPredictor class provides a wrapper over the week’s neural network implementation classes.