Mobility in Distributed Systems

Abstract

Mobile agents are programs with the additional capability to move between computers across a network connection. Movement implies that the running program that constitutes an agent moves from one system to another, taking with the agent the code that constitutes the agent as well as the state information of the agent. The movement of agents may be user-directed or self-directed (i.e. autonomous). In the case of user-directed movement, agents are configured with an itinerary that dictates the movement of the agents. In the case of self-directed movement, agents may move in order to better optimize their operation. Mobility may also be a combination of user- and self-directedness.

Mobile agents provide three basic capabilities: mobile code, mobile computation, and mobile state. These three capabilities are shown in the figure below. Each of the capabilities is an evolution of previously well-established notions in distributed and networked systems.

Mobile computation involves moving a computation from one system to another. This capability is an evolution of remote computation, which allows a system to tap into the computational resources of another system over a network connection. One of the original mechanisms for remote computation was Remote Procedure Call (RPC). Java Remote Method Invocation (RMI) is another example of remote computation as are servlets and stored procedures.

The difference between mobile and remote computation is that mobile computation supports network disconnection. In a traditional remote computation model, the system requesting the service (the client) must remain connected to the system providing the service (the server) for the duration of the remote computation operation. Additionally, depending on the interface exposed by the server, an interaction can require an arbitrary number of messages between client and server. If network connectivity is lost, the remote computation will become an orphaned computation that will either be terminated or whose results will be discarded. A mobile computation, on the other hand, is an autonomous entity. Once the computation moves from the first system (which may nominally be called the client) to the second system (the server), the computation continues to execute on the server even if the client becomes disconnected. The agent returns to the client with the results of the computation when (and if) the connectivity is recovered.

Mobile Code is the ability to move code from one system to another. The code may be either source code that is compiled or interpreted or binary code. Binary code may further be either machine dependent or be some intermediate, machine-independent form.

Mobile code is used in other contexts besides mobile agents. For example, system administrators use mobile code in order to remotely install or upgrade software on client systems. Similarly, a web browser uses mobile code to pull an applet or script to execute as part of a web page.

Code may be mobile in two different ways: push and pull. In the push model, the system sending the code originates the code transfer operation whereas in the pull model, the system receiving the code originates the code transfer operation. An example of the pull model is a Web browser downloading components such as applets or scripts. Remote installation is an example of the push model. Mobile agent systems use the push model of code mobility.

Pull mobility is often considered to be more secure and trustworthy because the host receiving the code is the one that requested the code. Usually, the origin of the request lies in some action carried out by a user of the system and hence pull mobility is superficially more secure. Push mobility on the other hand allows a system to send code to the receiving system at unexpected or unmonitored times. Hence push mobility is less trustworthy from a user’s point of view. In practice the overwhelming majority of security exploits encountered in distributed systems originates in careless user actions such as running mail attachments.

Mobile code allows systems to be extremely flexible. New capabilities can be downloaded to systems on the fly thereby dynamically adding features or upgrading existing features. Moreover, if capabilities can be downloaded on demand, temporarily unused capabilities can also be discarded. Swapping capabilities on an as-needed basis allows systems to support small memory constrained devices. Discarding capabilities after use can also help improve system security.

Mobile state is an evolution of state capture, which allows the execution state of a process to be captured. State capture has been traditionally used for checkpointing systems to protect against unexpected system failure. In the event of a failure, the execution of a process can be restarted from the last checkpointed state thereby not wasting time by starting from the very beginning. Checkpointing is thus very useful for long-running processes. Operating system research has investigate capturing entire process states, a variant of checkpointing, for load balancing purposes in the early 1980s, but that avenue of research proved to be a dead-end due to coarse granularity of process and semantics problem due to the impossibility of capturing operating system resources such as open file descriptors.

Mobile state allows the movement of the execution state of an agent to another system for continued execution. The key advantage provided by mobile state is that the execution of the agent does not need to restart after the agent moves to a new host. Instead, the execution continues at the very next instruction in the agent.

Not all mobile agent systems provide support for state mobility. The term strong mobility is used to describe systems that can capture and move execution state with the agent. Operationally, strong mobility guarantees that all variables will have identical values and the program counter will be at the same position. Weakly mobile-agent systems, on the other hand, usually support the capture of most of a program’s data, but restart the program from a predefined program point and thus require some programmer involvement at each migration. The advantage of strong mobility is that the result of migrating is well defined and easier to understand, but its disadvantage is that it is much more complex to implement efficiently.

The most important advantage provided by strong mobility is the ability to support external asynchronous migration requests (also known as forced mobility). This allows entities other than the agent (such as other system components, an administrator, or the owner) to request that an agent be moved. Forced mobility is useful for survivability, load-balancing, forced isolation, and replication for fault-tolerance, and is an important aspect of the agile computing approach.

For some applications, the technical issues of software mobility are dwarfed by security issues. Despite fits and starts, mobile agents herald emerging technological solutions. This special issue provides a good sampling of this frontier. It is our hope and desire that the software industry embrace the capabilities offered by mobile agents. In the future, we hope to see general purpose programming languages that support mobility as well as large scale development platforms and toolkits.

Niranjan Suri
Institute of Human and Machine Cognition

Henry Hexmoor
Sourthern Illinois University