Centro Tecnológico - Programa de Pós-Graduação em Informática

Av. Fernando Ferrari, s/n - Campus de Goiabeiras

CEP 29060-900 - Vitória - ES – Brasil

{tavares, fvarejao}@inf.ufes.br

Keywords: multi-agent system, network management, unix, intelligent agent, gaia methodology.

Currently, the computational environment of many companies is composed by a great variety of hardware components, software, operating systems, etc. Depending directly on their computational environment, there are several applications (systems) that companies use to support their main businesses. The stability and readiness of this computational environment have become more and more important so as to allow greater productivity and efficiency of companies. Slowness or even unavailability of some of the company most critical applications could be the result of any unstable or unavailable part of the computational environment, and can generate large financial damages [10] [12].

Analysis of the log of our experiment has shown that the multi-agent system avoided that important financial and production control systems were interrupted and provided disk space for executing routine backup procedures (which, otherwise, would not be performed).

This article has the following structure: section 2 introduces a general view of the area of network management and services. Section 3 presents a description of the multi-agent system. In section 4, we describe the experiment being performed. Finally, in section 5, the conclusions and future works are presented.

The network management main approach has been the IETF Management Structure (SNMP). The SNMP structure is based on a small group of concepts: agents or managed nodes (which represent elements of the network, as bridges, routers, switches, etc.), manager or network management system (which provides a set of operations to the network support team), managed information base - MIB (represents the information that can be manipulated by each agent), the management protocol - SNMP (standard way of extracting management information from the systems) and the proxy (allows the interaction of the SNMP structure with components that don't include this structure) [10] [12].

The SNMP structure is a centralized approach that uses distributed agents to collect manageable information. However, this approach has some problems when dealing with extensive and complex network models. The central manager may have some troubles for handling great amounts of manageable information and for managing a large number of geographically distributed sites [12] [16].

An alternative approach uses the "divide and conquer" strategy for promoting solutions based on distributed management and on delegation [12]. Several works use this approach, such as proposed by BRITES [2], GOLDSZMIDT [7] and OLIVEIRA [14]. In this approach, a main manager delegates the control to several distributed workstations and to more powerful agents. By promoting the collaboration among these agents, we have several management "islands". This procedure increases the reliability and failure tolerance of the network environment [12].

Another approach, considered by many researchers as the ideal solution for the management complex network management, is the paradigm of mobile agents [4] [10]. The term Mobile Agent refers to autonomous programs that can move through the network, from node to node, and assume the knowledge of an agent, i.e., acting as users or other entities [15]. Researches have been made in distributed network management and delegation [7], network services deliver [9], optimization of network traffic and failure tolerance of networks [11].

Whichever approach is adopted, it should aim to promote the pro-active concept in the network management. Several approaches have used this concept, such as, CHIU[5], DE FRANCESCHI [6] and HOOD [8].

The multi-agent approach is based on distributed management and on delegation. The agents are permanently acting in specific parts of the computational environment, being able to solve from simple routine problems (of their strict domain) to even more complex problems that involve communications and activities among several agents. Our approach was chosen based on the characteristics of the problem to be solved (the need for immediate actions instead of having to wait for human intervention), as well as on the intelligent agents characteristics (autonomy to take actions, always active and ready to take actions, capacity to act together with other agents, speed equivalent to the computational processors in which they are installed, etc.) [17].

The multi-agent system was developed using the GAIA methodology [18] for analysis and project of systems based on agents. The multi-agent system includes the following roles: ADMINISTRATOR, MONITOR, COORDINATOR and CONTROLLERS.

The computational environment represented in figure 1 is composed of several hardware components that host multi-agent systems. Inside these hardware components, there are software components and other components that the system may control. The monitor role is also part of the computational environment and is represented by the monitoring tools, as shown by the "black box" in figure 1.

Figure 1 - Global Architecture

Each multi-agent system has an agent that coordinates the other controlling agents. These agents are responsible for the communications among the controlling agents and also allow the interaction between the multi-agent system and the human experts. In addition, they receive the information about the problems detected by the monitoring tools and distribute them to the appropriate controlling agents.

The controlling agents are responsible for the specific control of services and parts of the computational environment. For instance, we may have safety controllers, database controllers, response time, disk space, cleaning, printing service, etc. Researches with similar architectures are found in the works of ABECK [1] and MARZO [13].

All agents have the same architecture. This feature allows the construction of a unique agent programming structure. The agent architecture is based on activities and rule bases. The agents should perform activities and apply their rule bases to accomplish them. The distinction between agents is expressed in their rule bases. Figure 2 shows that, if there is no longer any activity to be accomplished, the agents change their state to "sleeping". This state is disabled as soon as a new activity to be accomplished is sent to the agent or the sleeping time ceases.

Figure 2 - Agent Processing Architecture

Every agent selects the highest priority activity to perform first, as viewed in Figure 2. The fixed activities are executed after the "sleeping" state. The priorities sequence guarantees the compliance with the proposed hierarchical structure because an agent will always accomplish an activity generated by a human expert first, no matter how many fixed or communications activities it must have to accomplish in a given moment.

We developed a coordinator, a disk space controller, a processes controller, and a cleaning controller agent for solving this problem. The disk space controlling agent has the function of controlling the allocation of the computer disk space in order to maintain the environment stability. The processes controlling agent is responsible for controlling all the processes executed in the computer, analyzing CPU consumption, looping states, process ownership and the interdependencies among the processes. The cleaning controlling agent is responsible for controlling and optimizing the use of files inside the computer. It controls where the main files are and who is responsible for them. It also performs the cleaning of temporary files.

The strategy to obtain results in the experiment was divided in choosing an appropriate environment for the multi-agent system operation, choosing a problem with appropriate complexity degree, running the multi-agent system in the real environment during a period of three months and finally analyzing the results obtained.

The experiment development strategy consisted of choosing a part of the computational environment that was critical in terms of availability, involving a high level of complexity and demanding non trivial solutions and the participation of several controlling agents acting together to solve them. The Unix environment was the one chosen. The main applications assisting the metallurgical processes of CST use databases installed in Unix servers. The prototype was initially installed on eight Unix servers: two of development, one of endorsement and five of production (financial, supply, histories, control and supervision). The language chosen for the development of the agents was the language Korn Shell because it is native of the Unix environment (thus, computationally efficient) and allows the complete environment control.

The file system problem usually happens when a Unix environment process saves (or is still saving) large files or a large number of small files in the disk area. This problem solution consists of canceling the process creating the files, whenever it is still active, and then removing or moving those files to some free disk area in the same computer or in another one. That problem allows a good observation of the multi-agent system operation, because it demands the constant communication among the agents to be solved. Another important factor to the experiment is that the file system full problem often happens many times and at any time during a week, making possible real verification of the multi-agent system operation. Finally, the causes and solutions of the problem were very well defined, as well as the flow of communications and activities that the agents would have to execute.

Initially, the monitoring tool sends an activity for the coordinating agent, whose content informs where the area with problems is and the current rate of occupation. In the experiment, the monitoring tool was configured to verify all the areas of the computer every 3 minutes and to generate a problem for the multi-agent system, if some rate was above 95% of occupation.

The coordinating agent verifies the type of the activity to be executed and sends a message to the appropriate agent. In that case, the space disk controlling agent is the responsible for coping with that activity. In the experiment, the coordinating agent knowledge base was fulfilled with solutions of several types of problems.

The space disk controlling agent, when receiving the message, initially verifies if the problem continues to happen. Confirming that, it tries to identify the causes of the problem. Identifying the causes, it sends messages requiring cancellation of processes and cleaning of files to the coordinating agent, which, has the role of distributing the new activities to the appropriate agents. In the experiment, the causes were precisely identified. They differed from problem to problem. The system was able to be sometimes originated by one or more files with an or more processes. Problems found in the identification were due to the current lack of rights for investigation in certain directories or because the problem had already finished when the agent began the identification. In that case, the agent concluded that there was nothing to do.

The coordinating agent received activities again, this time originated by the space disk controlling agent and identifies which agents would be the responsible for those.

The processes controlling agent and the cleaning controlling agent had in your knowledge bases, relationships of countermands for cancellation of certain processes as well as removal certain files. In the experiment, when those countermands were identified the specific agent could not solve the activity and it returned that information for the coordinating agent that directs the impossibility for the space disk controlling agent. In this case, the space disk controlling agent documents the impossibility of solving the problem and informs the specialists about the found countermands. If there is no countermands, the processes controlling agent would usually cancel the requested processes and the cleaning controlling agent would remove the appropriate files. Until we consolidate the acceptance of the system, the agents inform the specialists what they would have done to solve the problem, as well as informing the responsible for the affected systems of those actions. The analysis of the logs of the actions taken until now were very positive. There was not any indication of disastrous actions taken over the environment.

Finally, the controlling agents send a message to the coordinating agent indicating the end of the activities execution. The coordinating agent sends this information to the space disk controlling agent.

In the table 1, we see a summary of the obtained results.

Table 1: Results Obtained on the Multi-agent System Experiment

The reasons of the non resolution of problems for the multi-agent system were caused by the lack of rights for investigating certain directories and also because the problem had already been finished at the moment of the analysis. The results obtained by the experiment have been very positive. Most of the problems were correctly identified and also were their causes.

This work shows how an Artificial Intelligence based technology, not frequently used in the routine working environment of companies, can be of great value for keeping available the company computational environment.

The multi-agent system may avoid to stop parts of the industrial production, failures in performing financial operations that may provoke fines, human expert time spent solving complex problems that began simple, loss of time from users, customers having their purchases impeded, accidents with our without human loss, etc. The good results obtained by the multi-agent system have motivated claims from the human experts to include new and more complex problems to be solved. New knowledge is being included to the agent rule bases and new agents are being constructed to attend these demands.

In the short term, we intend to add more controlling agents to the Unix environment to evaluate the growing difficulties of an increasingly complex agent structure acting together. The next controllers will be the database ones (a type of omnipresent DBA), the safety ones (to control the access to the computers, invasions attempts, password control, etc.), the response time (to maintain the environment with suitable performance rates) and other specific services controllers. It is also our intention to develop in the medium term agents using a language compatible with other environments that can be controlled (NT, Netware, WEB, etc.). In the long term, we intend to endow the agents with a type of learning capacity, through the development of "curious" or "questioning" agents which register situations in which the multi-agent systems have failed. These agents will keep questioning the human experts about how these agents were solved. Another type of learning happens when the agents themselves inform the agents of the same type located in other hardware about alterations in their knowledge base. With the growing complexity of the knowledge bases, it would be ideal to have a way to make the agents themselves optimize those bases in order to get better solutions than the ones previously used. In this context, the agents will be able to teach the human experts.