[TOC] [Prev] [Next]

System Architecture

Note: The RMI documentation on this site is temporarily unavailable. Links to all RMI pages will be redirected shortly.



The RMI system consists of three layers: the stub/skeleton layer, the remote reference layer, and the transport layer. The boundary at each layer is defined by a specific interface and protocol; each layer, therefore, is independent of the next and can be replaced by an alternate implementation without affecting the other layers in the system. For example, the current transport implementation is TCP-based (using Java sockets), but a transport based on UDP could be substituted.

To accomplish transparent transmission of objects from one address space to another, the technique of object serialization (designed specifically for the Java language) is used. Object serialization is described in this chapter only with regard to its use for marshaling primitives and objects. For complete details, see the Object Serialization Specification.

Another technique, called dynamic stub loading, is used to support client-side stubs which implement the same set of remote interfaces as a remote object itself. This technique, used when a stub of the exact type is not already available to the client, allows a client to use the Java language's built-in operators for casting and type-checking.

Architectural Overview

The RMI system consists of three layers:

The application layer sits on top of the RMI system. The relationship between the layers is shown in the following figure.

A remote method invocation from a client to a remote server object travels down through the layers of the RMI system to the client-side transport, then up through the server-side transport to the server.

A client invoking a method on a remote server object actually makes use of a stub or proxy for the remote object as a conduit to the remote object. A client- held reference to a remote object is a reference to a local stub. This stub is an implementation of the remote interfaces of the remote object and forwards invocation requests to that server object via the remote reference layer. Stubs are generated using the rmic compiler.

The remote reference layer is responsible for carrying out the semantics of the invocation. For example, the remote reference layer is responsible for determining whether the server is a single object or is a replicated object requiring communications with multiple locations. Each remote object implementation chooses its own remote reference semantics-whether the server is a single object or is a replicated object requiring communications with its replicas.

Also handled by the remote reference layer are the reference semantics for the server. The remote reference layer, for example, abstracts the different ways of referring to objects that are implemented in (a) servers that are always running on some machine, and (b) servers that are run only when some method invocation is made on them (activation). At the layers above the remote reference layer, these differences are not seen.

The transport layer is responsible for connection setup, connection management, and keeping track of and dispatching to remote objects (the targets of remote calls) residing in the transport's address space.

In order to dispatch to a remote object, the transport forwards the remote call up to the remote reference layer. The remote reference layer handles any server-side behavior that needs to occur before handing off the request to the server-side skeleton. The skeleton for a remote object makes an up call to the remote object implementation which carries out the actual method call.

The return value of a call is sent back through the skeleton, remote reference layer, and transport on the server side, and then up through the transport, remote reference layer, and stub on the client side.

The Stub/Skeleton Layer

The stub/skeleton layer is the interface between the application layer and the rest of the RMI system. This layer does not deal with specifics of any transport, but transmits data to the remote reference layer via the abstraction of marshal streams. Marshal streams employ a mechanism called object serialization which enables Java objects to be transmitted between address spaces. Objects transmitted using the object serialization system are passed by copy to the remote address space, unless they are remote objects, in which case they are passed by reference.

A stub for a remote object is the client-side proxy for the remote object. Such a stub implements all the interfaces that are supported by the remote object implementation. A client-side stub is responsible for:

A skeleton for a remote object is a server-side entity that contains a method which dispatches calls to the actual remote object implementation. The skeleton is responsible for:

The appropriate stub and skeleton classes are determined at run time and are dynamically loaded as needed, as described in Dynamic Class Loading. Stubs and skeletons are generated using the rmic compiler.

The Remote Reference Layer

The remote reference layer deals with the lower-level transport interface. This layer is also responsible for carrying out a specific remote reference protocol which is independent of the client stubs and server skeletons.

Each remote object implementation chooses its own remote reference subclass that operates on its behalf. Various invocation protocols can be carried out at this layer. Examples are:

The remote reference layer has two cooperating components: the client-side and the server-side components. The client-side component contains information specific to the remote server (or servers, if the remote reference is to a replicated object) and communicates via the transport to the server-side component. During each method invocation, the client and server-side components perform the specific remote reference semantics. For example, if a remote object is part of a replicated object, the client-side component can forward the invocation to each replica rather than just a single remote object.

In a corresponding manner, the server-side component implements the specific remote reference semantics prior to delivering a remote method invocation to the skeleton. This component, for example, would handle ensuring atomic multicast delivery by communicating with other servers in a replica group (note that multicast delivery is not part of the JDK 1.1 release of RMI).

The remote reference layer transmits data to the transport layer via the abstraction of a stream-oriented connection. The transport takes care of the implementation details of connections. Although connections present a streams-based interface, a connectionless transport can be implemented beneath the abstraction.

The Transport Layer

In general, the transport layer of the RMI system is responsible for:

The concrete representation of a remote object reference consists of an endpoint and an object identifier. This representation is called a live reference. Given a live reference for a remote object, a transport can use the endpoint to set up a connection to the address space in which the remote object resides. On the server side, the transport uses the object identifier to look up the target of the remote call.

The transport for the RMI system consists of four basic abstractions:

A transport defines what the concrete representation of an endpoint is, so multiple transport implementations may exist. The design and implementation also supports multiple transports per address space, so both TCP and UDP can be supported in the same virtual machine. Note that the RMI transport interfaces are only available to the virtual machine implementation and are not available directly to the application.

Thread Usage in Remote Method Invocations

A method dispatched by the RMI runtime to a remote object implementation (a server) may or may not execute in a separate thread. Some calls originating from the same client virtual machine will execute in the same thread; some will execute in different threads. Calls originating from different client virtual machines will execute in different threads. Other than this last case of different client virtual machines, the RMI runtime makes no guarantees with respect to mapping remote object invocations to threads.

Garbage Collection of Remote Objects

In a distributed system, just as in the local system, it is desirable to automatically delete those remote objects that are no longer referenced by any client. This frees the programmer from needing to keep track of the remote objects clients so that it can terminate appropriately. RMI uses a reference- counting garbage collection algorithm similar to Modula-3's Network Objects. (See "Network Objects" by Birrell, Nelson, and Owicki, Digital Equipment Corporation Systems Research Center Technical Report 115, 1994.)

To accomplish reference-counting garbage collection, the RMI runtime keeps track of all live references within each Java virtual machine. When a live reference enters a Java virtual machine, its reference count is incremented. The first reference to an object sends a "referenced" message to the server for the object. As live references are found to be unreferenced in the local virtual machine, their finalization decrements the count. When the last reference has been discarded, an unreferenced message is sent to the server. Many subtleties exist in the protocol; most of these are related to maintaining the ordering of referenced and unreferenced messages in order to ensure that the object is not prematurely collected.

When a remote object is not referenced by any client, the RMI runtime refers to it using a weak reference. The weak reference allows the Java virtual machine's garbage collector to discard the object if no other local references to the object exist. The distributed garbage collection algorithm interacts with the local Java virtual machine's garbage collector in the usual ways by holding normal or weak references to objects. As in the normal object life-cycle finalize will be called after the garbage collector determines that no more references to the object exist.

As long as a local reference to a remote object exists, it cannot be garbage- collected and it can be passed in remote calls or returned to clients. Passing a remote object adds the identifier for the virtual machine to which it was passed to the referenced set. A remote object needing unreferenced notification must implement the java.rmi.server.Unreferenced interface. When those references no longer exist, the unreferenced method will be invoked. unreferenced is called when the set of references is found to be empty so it might be called more than once. Remote objects are only collected when no more references, either local or remote, still exist.

Note that if a network partition exists between a client and a remote server object, it is possible that premature collection of the remote object will occur (since the transport might believe that the client crashed). Because of the possibility of premature collection, remote references cannot guarantee referential integrity; in other words, it is always possible that a remote reference may in fact not refer to an existing object. An attempt to use such a reference will generate a RemoteException which must be handled by the application.

Dynamic Class Loading

In RPC (remote procedure call) systems, client-side stub code must be generated and linked into a client before a remote procedure call can be done. This code can be either statically linked into the client or linked in at runtime via dynamic linking with libraries available locally or over a network file system. In the case of either static or dynamic linking, the specific code to handle an RPC must be available to the client machine in compiled form.

RMI generalizes this technique, using a mechanism called dynamic class loading to load at runtime (in the Java language's architecture neutral bytecode format) the classes required to handle method invocations on a remote object. These classes are:

This section describes:

In addition to class loaders, dynamic class loading employs two other mechanisms: the object serialization system to transmit classes over the wire, and a security manager to check the classes that are loaded. The object serialization system is discussed in the Object Serialization Specification. Security issues are discussed in Security.

How a Class Loader is Chosen

In Java, the class loader that initially loads a Java class is subsequently used to load all the interfaces and classes that are used directly in the class:

For objects passed as parameters or return values (the second case above), the URL that is encoded in the stream for an object's class is determined as follows:

Thus, if a class was loaded from CLASSPATH, the codebase URL will be used to annotate that class in the stream if that class is used in an RMI call.

The application can be configured with the property java.rmi.server.useCodebaseOnly, which disables the loading of classes from network hosts and forces classes to be loaded only from the locally defined codebase. If the required class cannot be loaded, the method invocation will fail with an exception.

Bootstrapping the Client

For the RMI runtime to be able to download all the classes and interfaces needed by a client application, a bootstrapping client program is required which forces the use of a class loader (such as RMI's class loader) instead of the default class loader. The bootstrapping program needs to:

For example:

import java.rmi.RMISecurityManager;
import java.rmi.server.RMIClassLoader;

public class LoadClient
	public static void main()
		System.setSecurityManager(new RMISecurityManager());
		try {
			Class cl = RMIClassLoader.loadClass("myclient");
			Runnable client = (Runnable)cl.newInstance();
		} catch (Exception e) {
			System.out.println("Exception: " + e.getMessage());
In order for this code to work, you need to specify the java.rmi.server.codebase property when you run the bootstrapping program so that the loadClass method will use this URL to load the class. For example:

java -Djava.rmi.server.codebase=http://host/rmiclasses/ LoadClient
Instead of relying on the property, you can supply your own URL:

Class cl = RMIClassLoader.loadClass(url, "myclient");
Once the client is started and has control, all classes needed by the client will be loaded from the specified URL. This bootstrapping technique is exactly the same technique Java uses to force the AppletClassLoader to download the same classes used in an applet.

Without this bootstrapping technique, all the classes directly referenced in the client code must be available through the local CLASSPATH on the client, and the only Java classes that can be loaded by the RMIClassLoader over the net are classes that are not referred to directly in the client program; these classes are stubs, skeletons, and the extended classes of arguments and return values to remote method invocations.


In Java, when a class loader loads classes from the local CLASSPATH, those classes are considered trustworthy and are not restricted by a security manager. However, when the RMIClassLoader attempts to load classes from the network, there must be a security manager in place or an exception is thrown.

The security manger must be started as the first action of a Java program so that it can regulate subsequent actions. The security manager ensures that loaded classes adhere to the standard Java safety guarantees, for example that classes are loaded from "trusted" sources (such as the applet host) and do not attempt to access sensitive functions. A complete description of the restrictions imposed by security managers can be found in the documentation for the AppletSecurity class and the RMISecurityManager class.

Applets are always subject to the restrictions imposed by the AppletSecurity class. This security manager ensures that classes are loaded only from the applet host or its designated codebase hosts. This requires that applet developers install the appropriate classes on the applet host.

Applications must either define their own security manager or use the restrictive RMISecurityManager. If no security manager is in place, an application cannot load classes from network sources.

A client or server program is usually implemented by classes loaded from the local system and therefore is not subject to the restrictions of the security manager. If however, the client program itself is downloaded from the network using the technique described in Bootstrapping the Client, then the client program is subject to the restrictions of the security manager.

Once a class is loaded by the RMIClassLoader, any classes used directly by that class are also loaded by the RMIClassLoader and thus are subject to the security manager restrictions.

Even if a security manager is in place, setting the property java.rmi.server.useCodebaseOnly to true prevents the downloading of a class from the URL embedded in the stream with a serialized object (classes can still be loaded from the locally-defined java.rmi.server.codebase). The java.rmi.server.useCodebaseOnly property can be specified on both the client and the server, but is not applicable for applets.

If an application defines its own security manager which disallows the creation of a class loader, classes will be loaded using the default Class.forName mechanism. Thus, a server may define its own policies via the security manager and class loader, and the RMI system will operate within those policies.

The java.lang.SecurityManager abstract class, from which all security managers are extended, does not regulate resource consumption. Therefore, the current RMISecurityManager has no mechanisms available to prevent classes loaded from abusing resources. As new security manager mechanisms are developed, RMI will use them.

Configuration Scenarios

The RMI system supports many different scenarios. Servers can be configured in an open or closed fashion. Applets can use RMI to invoke methods on objects supported on servers. If an applet creates and passes a remote object to the server, the server can use RMI to make a callback to the remote object. Java applications can use RMI either in client-server mode or from peer to peer. This section highlights the issues surrounding these configurations.


The typical closed-system scenario has the server configured to load no classes. The services it provides are defined by remote interfaces that are all local to the server machine. The server has no security manager and will not load classes even if clients send along the URL. If clients send remote objects for which the server does not have stub classes, those method invocations will fail when the request is unmarshaled, and the client will receive an exception.

The more open server system will define its java.rmi.server.codebase so that classes for the remote objects it exports can be loaded by clients, and so that the server can load classes when needed for remote objects supplied by clients. The server will have both a security manager and RMI class loader which protect the server. A somewhat more cautious server can use the property java.rmi.server.useCodebaseOnly to disable the loading of classes from client-supplied URLs.


Typically, the classes needed will be supplied by an HTTP server or by an FTP server as referenced in URL's embedded in the HTML page containing the applet. The RMI-based service(s) used by the applet must be on the server from which the applet was downloaded, because an applet can only make network connections to the host from which it was loaded.

For example, the normal applet scenario uses a single host for the HTTP server providing the HTML page, the applet code, the RMI services, and the bootstrap Registry. In this scenario, all the stub, skeleton, and supporting classes are loaded from the HTTP server. All of the remote objects provided by the RMI service and passed to the applet (which may pass them back to the server) will be for classes that the RMI service already knows about. In this case, the RMI service is very secure because it loads no classes from the network.


Applications written in the Java language, unlike applets, can connect to any host; so Java applications have more options for configuring the sources of classes and where RMI based services run. Typically, a single HTTP server will be used to supply remote classes, while the RMI-based applications themselves are distributed around the network on servers or running on user's desktops

If an application is loaded locally, then the classes used directly in that program must also be available locally. In this scenario, the only classes that can be downloaded from a network source are the classes of remote interfaces, stub classes, and the extended classes of arguments and return values to remote method invocations.

If an application is not loaded from a local directory, but is loaded from a network source using the bootstrapping mechanism described in Bootstrapping the Client, then all classes used by the application can be downloaded from the same network source.

To enable downloading from a network source, each remote object server must be configured with the java.rmi.server.codebase property which specifies where application classes and generated stubs/skeletons reside. When the codebase property is specified, the RMI system embeds the URL of a class in the serialized form of the class.

Even if a serialized object's class is annotated with the URL from which the class can be downloaded, a client or peer will still load classes locally if they are available.

RMI Through Firewalls Via Proxies

The RMI transport layer normally attempts to open direct sockets to hosts on the Internet. Many intranets, however, have firewalls which do not allow this. The default RMI transport, therefore, provides two alternate HTTP-based mechanisms which enable a client behind a firewall to invoke a method on a remote object which resides outside the firewall.

How an RMI Call is Packaged within the HTTP Protocol

To get outside a firewall, the transport layer embeds an RMI call within the firewall-trusted HTTP protocol. The RMI call data is sent outside as the body of an HTTP POST request, and the return information is sent back in the body of the HTTP response. The transport layer will formulate the POST request in one of two ways:

  1. If the firewall proxy will forward an HTTP request directed to an arbitrary port on the host machine, then it is forwarded directly to the port on which the RMI server is listening. The default RMI transport layer on the target machine is listening with a server socket that is capable of understanding and decoding RMI calls inside POST requests.
  2. If the firewall proxy will only forward HTTP requests directed to certain well-known HTTP ports, then the call will be forwarded to the HTTP server listening on port 80 of the host machine, and a CGI script will be executed to forward the call to the target RMI server port on the same machine.

The Default Socket Factory

The RMI transport extends the java.rmi.server.RMISocketFactory class to provide a default implementation of a socket factory which is the resource- provider for client and server sockets. This default socket factory creates sockets that transparently provide the firewall tunnelling mechanism as follows:

Client-side sockets, with this default behavior, are provided by the factory's java.rmi.server.RMISocketFactory.createSocket method. Server- side sockets with this default behavior are provided by the factory's java.rmi.server.RMISocketFactory.createServerSocket method.

Configuring the Client

There is no special configuration necessary to enable the client to send RMI calls through a firewall.

The client can, however, disable the packaging of RMI calls as HTTP requests by setting the java.rmi.server.disableHttp property to equal the boolean value true.

Configuring the Server

The host name should not be specified as the host's IP address, because some firewall proxies will not forward to such a host name.

  1. In order for a client outside the server host's domain to be able to invoke methods on a server's remote objects, the client must be able to find the server. To do this, the remote references that the server exports must contain the fully-qualified name of the server host. Depending on the server's platform and network environment, this information may or may not be available to the Java virtual machine on which the server is running. If it is not available, the host's fully qualified name must be specified with the property java.rmi.server.hostname when starting the server.

    For example, use this command to start the RMI server class ServerImpl on the machine chatsubo.javasoft.com:

       java -Djava.rmi.server.hostname=chatsubo.javasoft.com ServerImpl
  2. If the server will not support RMI clients behind firewalls that can forward to arbitrary ports, use this configuration:
    1. An HTTP server is listening on port 80.
    2. A CGI script is located at the aliased URL path /cgi-bin/java-rmi. This script:

Performance Issues and Limitations

Calls transmitted via HTTP requests are at least an order of magnitude slower that those sent through direct sockets, without taking proxy forwarding delays into consideration.

Because HTTP requests can only be initiated in one direction through a firewall, a client cannot export its own remote objects outside the firewall, because a host outside the firewall cannot initiate a method invocation back on the client.

[TOC] [Prev] [Next]

Copyright © 1996, 1997 Sun Microsystems, Inc. All rights reserved.