The Architecture of Aglets

The Inner Workings of an Agent Technology

by Bill Venners
April 15, 1997

First published in JavaWorld, April 1997
Mobile agents have been around for many years, but they haven't yet entered the mainstream. This article takes a look at aglets, a mobile-agent technology built on top of Java.

Welcome to another edition of Under The Hood. Up to now, this column has focused on the inner workings of the Java virtual machine (JVM). I've written overviews of the JVM, the class file, and garbage collection, and have covered most of the JVM's bytecode instruction set. I have one final bytecode article coming in June, but this month I am going to begin expanding the column's scope. In the future, I plan to explore a broader array of topics. Each month I will focus on a particular aspect or application of Java technology, explain "how it works," and analyze what it means to Java developers.

This month's article looks at aglets, an innovation developed by IBM Japan.

Aglets: Not just for shoelaces anymore
According to Webster's Ninth New Collegiate Dictionary, an aglet is:

  • aglet 1: the plain or ornamental tag covering the ends of a lace
  • aglet 2: any of various ornamental studs, cords, or pins worn on clothing

In other words, aglets are those little plastic tubes on the ends of your shoelaces. Now, however, there is a new definition of the word aglet: a Java-based autonomous software agent.

As used here, a software agent is a program that can halt itself, ship itself to another computer on the network, and continue execution at the new computer. An agent doesn't restart execution from the beginning at the new computer; it continues where it left off. For example, imagine an agent that increments a counter starting with zero. If that agent counts from zero to ten, then halts and ships itself to another computer, it will not start counting again at zero. It will continue counting starting with ten, because that was where it left off when it halted at its previous computer.

Agents are autonomous because they decide where they will go and what they will do. They control their lifetimes. They can receive requests from external sources, such as other agents, but each individual agent decides whether or not to comply with external requests. Also, agents can decide to perform actions, such as travel across a network to a new computer, independent of any external request.

Aglets versus applets
The Java aglet extends the model of network-mobile code made famous by Java applets. Like an applet, the class files for an aglet can migrate across a network. But unlike applets, when an aglet migrates it also carries its state. An applet is code that can move across a network from a server to a client. An aglet is a running Java program (code and state) that can move from one host to another on a network. In addition, because an aglet carries its state wherever it goes, it can travel sequentially to many destinations on a network, including eventually returning back to its original host.

A Java aglet is similar to an applet in that it runs as a thread (or multiple threads) inside the context of a host Java application. To run applets, a Web browser fires off a Java application to host any applets it may encounter as the user browses from page to page. That application installs a security manager to enforce restrictions on the activities of any untrusted applets. To download an applet's class files, the application creates class loaders that know how to request class files from an HTTP server.

Likewise, an aglet requires a host Java application, an "aglet host," to be running on a computer before it can visit that computer. When aglets travel across a network, they migrate from one aglet host to another. Each aglet host installs a security manager to enforce restrictions on the activities of untrusted aglets. Hosts upload aglets through class loaders that know how to retrieve the class files and state of an aglet from a remote aglet host.

The aglet lifestyle
An aglet can experience many events in its life. It can be:

Created: a brand new aglet is born -- its state is initialized, its main thread starts executing
Cloned: a twin aglet is born -- the current state of the original is duplicated in the clone
Dispatched: an aglet travels to a new host -- the state goes with it
Retracted: an aglet, previously dispatched, is brought back from a remote host -- its state comes back with it
Deactivated: an aglet is put to sleep -- its state is stored on a disk somewhere
Activated: a deactivated aglet is brought back to life -- its state is restored from disk
Disposed of: an aglet dies -- its state is lost forever

Note that every activity besides creation and disposal involve either duplication, transmission across a network, or persistent storage of the aglet's state. Each of these activities uses the same process to get the state out of an aglet: serialization.

Serializing the state...
Aglet hosts use object serialization, available in JDK 1.1 or with the RMI (remote method invocation) add-on to JDK 1.0.2, to export the state of an aglet object to a stream of bytes. Through this process, the aglet object and the tree of serializable objects reachable from it, are written to a stream. An object is serializable if it implements either the Serializable or the Externalizable interface. In a reverse process, the state of the aglet can be reconstructed from the stream of bytes. Serialization allows an image of the heap (the heap's state) to be exported to a byte stream (such as a file) and then reconstructed from that byte stream.

...but not all of the state
The state of the execution stacks and program counters of the threads owned by the aglet are not serialized. Object serialization touches only data on the heap, not the stacks or the program counters. Thus when an aglet is dispatched, cloned, or deactivated, any relevant state sitting on any stack of a running aglet, as well as the current program counter for any thread, is lost.

In theory, a software agent should be able to migrate with all its state: heap, execution stack, and registers. Some will likely consider the inability of aglets to do this as a flaw in the aglet's implementation of mobile-agent theory. This feature of aglets arises out of the architecture of the JVM, which doesn't allow a program to directly access and manipulate execution stacks. This is part of the JVM's built-in security model. Unless there is a change to the JVM, aglets and any other mobile Java-based agent will be unable to carry the state of their execution stacks with them as they migrate.

Before it is serialized, an aglet must place on the heap everything it will need to know to be resurrected properly as a newly activated aglet, a freshly dispatched aglet, or a clone. It can't leave any of this information on the stack, because the stacks won't be reproduced in the aglet's new life. As a result, the aglet host informs an aglet that it is about to be serialized so that the aglet can prepare itself. When the aglet is informed of an impending serialization, it must place onto the heap any information it will need to continue its execution properly when it is resurrected.

>From a practical standpoint, the inability of an aglet to migrate with its execution stacks is not an unreasonable limitation. It simply forces you to think a certain way when you write aglets. You can look at an aglet as a finite state machine with the heap as the sole repository of the machine's state. If at any point in an aglet's life you can know what state it is in by looking at its heap, then it can be serialized at any time. If not, then you must have a way to record sufficient information on the heap just prior to serialization such that you can continue properly when the aglet is resurrected.

Also, even though the inability to serialize execution stacks necessitates giving aglets a warning prior to serialization, such warnings probably are a good idea anyway. It is difficult to think of a case in which an aglet wouldn't want to know it was about to be serialized and why. It may need to finish some incomplete process before allowing the serialization, or it may want to refuse the action that requires the serialization. For example, if an agent is told it is about to be serialized and dispatched to an aglet host in Silicon Valley, it may refuse and decide instead to dispatch itself to a host on an island in the South Pacific.

How to write an aglet
The process of writing an aglet is in many ways similar to the process of writing an applet. To create an applet, you subclass class Applet. To initialize an applet, you override the init() method, the starting point for any applet. You can use init() to build the user interface of the applet. If you wish, you can fire off other threads from init(). If you do this, you also may override stop() and start() to stop and restart your threads when the browser leaves and returns to the Web page. If you don't create any threads in init(), your applet likely will get at least one thread just because class Applet descends from class Panel. The AWT user-interface library of which Panel is a part will provide whatever threads are needed to run the user interface you create in init().

The aglet development and run-time environments provide a library of Java classes that support the creation and running of aglets. To create an aglet, you must subclass class Aglet, which includes several methods you can override to customize the behavior of your aglet. The aglet's counterpart to the init() method of applets is the onCreation() method. To initialize an aglet, you override onCreation(). The onCreation() method is invoked only once in an aglet's lifetime and should be used only for initialization.

The aglet also has a run() method, which represents the entry point for the aglet's main thread. This is similar to the main() method of a Java application, except that run() is invoked each time an aglet arrives at a new aglet host. For example, if you designed a CatAglet that visits nine different aglet hosts looking for MouseAglets, onCreation() would be invoked only once, when the CatAglet was first instantiated at its first host. Once onCreation() completed, run() would be invoked. Each time the CatAglet arrived at a new host, a method called onArrival() would be invoked to perform any initialization. Once onArrival() completed, run() would be invoked to get the aglet started again at the new host.

Starting run() again each time an aglet is brought to life illustrates the inability of aglets to transmit the state of their execution stacks. For example, imagine a HealthyAglet whose run() method periodically invokes a method named walk(). If, as it is walking, the HealthyAglet is serialized and transmitted to another host, it wouldn't by default continue executing where it left off in walk(). It would start over again at the beginning of run(). Thus, when the aglet is informed that it is about to be serialized, it would need to record on the heap that it is walking -- perhaps in an instance variable of HealthyAglet. That instance variable would be serialized and would migrate with the aglet. When run() is invoked to start the aglet's new life, the run() method would check the instance variable, see it was walking beforehand, and call walk().

The callback model
Before any major event in an aglet's life, a "callback" method is invoked to allow the aglet to prepare for (or refuse to partake in) the event. This is how an aglet learns that it is about to be serialized. For example, before an aglet is dispatched to a new location, the aglet's onDispatch() is invoked. This method indicates to an aglet that it is about to be sent to a new host, the URL of which is specified as a parameter to onDispatch(). In the body of onDispatch(), the aglet must decide whether or not to go. If the aglet decides it doesn't want to go, it throws an exception. If it decides to go, it must complete any unfinished business and prepare its state for serialization. When it returns from onDispatch(), its state will be serialized and all its threads terminated. The class files and serialized state will then be sent to the new host, where the aglet will be resurrected.

The method onDispatch() is a "callback" method because the aglet host invokes it some time after another method, dispatch(), is invoked. An aglet can invoke dispatch() on itself or on another aglet. This callback model for aglets is similar to that of windowing user interfaces. To repaint an AWT component, for example, you invoke the component's repaint() method. At some point later, the system calls back the component's update() method, which in turn calls paint().

The Aglet class defines these five callback methods, which you can override to customize the behavior of your aglet:

onCloning() -- called before a clone operation
onDispatch() -- called before a dispatch
onReverting() -- called before a retraction
onDeactivating() -- called before a deactivation
onDisposing() -- called before a dispose operation (Unlike real life, an aglet can throw an exception if it doesn't want to die.)

For each of these processes, the Aglet class has a corresponding method that triggers the action: clone(), dispatch(), retract(), deactivate(), and dispose(). Some time after these are called, the aglet host will invoke the appropriate callback method.

Each time an aglet begins execution at a host, the host invokes an initialization method on the aglet. When the initialization method returns, the host invokes run(). Depending on the event that precipitated the aglet's new life, the aglet host will choose to invoke one of these four initialization methods:

onCreation() -- called the first time an aglet springs to life
onClone() -- called on a clone after a clone operation
onArrival() -- called after a dispatch or a retraction
onActivation() -- called after an activation

Interaction between aglet and host
An aglet interacts with its environment (its aglet host) through an AgletContext object. An aglet can obtain a handle to its context by invoking getAgletContext(), a method it inherits from base class Aglet. The aglet context has methods such as createAglet() and retractAglet(), which allow an aglet to add new aglets (or get an old aglet back) to its local host.

Interaction between aglets
To interact with each other, aglets do not normally invoke each other's methods directly. Instead they go through AgletProxy objects, which serve as aglet representatives. For example, if a BossAglet wishes to make a request of an EmployeeAglet, the BossAglet obtains a handle to a proxy object that "represents" the EmployeeAglet. The BossAglet then makes a request by invoking a method in the EmployeeAglet's proxy, which in turn forwards the request to the actual EmployeeAglet.

The AgletProxy class contains methods that allow aglets to request other aglets to take actions, such as dispatch(), clone(), deactivate(), and dispose(). The aglet that has been requested to take an action can comply, refuse to comply, or decide to comply later.

The proxy also allows an aglet to send a message, either synchronously or asynchronously, to another aglet. A Message object is supplied for this purpose; it carries a String to indicate the kind of message plus one other optional piece of data, either a String or one of Java's primitive types. To send a message you create a Message object and pass it as a parameter to the sendMessage() or sendAsynchMessage() method of the proxy object.

An aglet must go through a proxy object to interact with an aglet, even if both aglets are in the same aglet host. The reason aglets aren't allowed to directly interact with one another is that the aglet's callback and initialization methods are public. These methods should be invoked only by the aglet host, but if an aglet could get a handle to another aglet, it could invoke that aglet's callback or initialization methods. An aglet could become very confused if another aglet inadvertently or maliciously invoked these methods directly.

The aglet being represented by a proxy might be local or remote, but the proxy object is always local. For example, if a BossAglet in Silicon Valley wants to communicate with an EmployeeAglet on a South Pacific island, the BossAglet gets a local AgletProxy object, which represents the remote EmployeeAglet. The BossAglet merely invokes methods in the local proxy, which in turn communicates across the network to the EmployeeAglet. Only aglets, not proxies, migrate across the network. A proxy communicates with a remote aglet that it represents by sending data across the network.

You get a proxy to an aglet in one of three ways, each of which involves invoking a method in the context object:

  1. By creating the aglet in the first place with createAglet(). (This returns a proxy object.)
  2. By searching through an enumeration of local proxies returned by getAgletProxies().
  3. By supplying an "aglet identifier" and, if remote, an aglet location as parameters to getAgletProxy(). (Every aglet, upon creation or cloning, is assigned a globally unique aglet identifier.)

Mobile-agent systems, such as aglets, require high levels of security, because they represent yet another way to transmit a malicious program. Before aglets can be used in practice, there must be an infrastructure of aglet hosts that prevent untrusted aglets from doing damage but provide trusted aglets with useful access to the host's resources. Security is amply provided for in Java's intrinsic architecture and in the extra security features of JDK 1.1, but as with applets, some attacks (such as denial of service by allocating memory until the host crashes) are still possible. Currently, the aglet hosts from IBM (named Tahiti and Fiji) place very severe security restrictions on the activities of any aglet that didn't originate locally.

Next month
Will aglets become as ubiquitous as their plastic cousins, which quietly perch on the ends of everyone's shoelaces? Aglets represent a good example of innovation on top of Java's network-oriented architecture, but what new benefits do they offer developers and end users that client/server, applets, and servlets don't already offer? In next month's Under The Hood, I will analyze the real-world utility of mobile agents in general and aglets in particular.


This article was first published under the name Under the Hood: The architecture of aglets in JavaWorld, a division of Web Publishing, Inc., April 1997.

Talk back!

Have an opinion? Readers have already posted 1 comment about this article. Why not add yours?

About the author

Bill Venners has been writing software professionally for 12 years. Based in Silicon Valley, he provides software consulting and training services under the name Artima Software Company. Over the years he has developed software for the consumer electronics, education, semiconductor, and life insurance industries. He has programmed in many languages on many platforms: assembly language on various microprocessors, C on Unix, C++ on Windows, Java on the Web. He is author of the book: Inside the Java Virtual Machine, published by McGraw-Hill.