Tcl DP [#!TclDP!#] is the most popular extension for distributed scripting. Our first point of comparison is with this system. In contrast to Tcl DP that is RPC oriented, our system is intended as a platform to script distributed event-oriented applications. We rely on one-way messages to support this. In Tcl DP, messages are round-trip and the sender cannot proceed until the recipient has completed processing. Our system can also support synchronous (round-trip) messages where the sender blocks until the append handler at the target completes execution and hence we can do the kinds of things Tcl DP is aimed at doing. However, we expect most applications built using our system to be one-way message oriented. It is interesting to note that in our system, we can move the server in response to an RPC before the reply comes back to the client.
Our framework and toolkit is related to several other systems that support mobility. In contrast to other research in Mobile Agents, our approach has been to treat mobility and Mobile Agent technology as an enhancement to distributed scripting rather than as a means of supporting disconnected operations. Consequently, we have concentrated on typical distributed systems issues such as location tracking, message passing and failure handling. This distinguishes and separates our work from the other work in this area. Tcl provides an ideal platform for building mobile agent systems and there have been a few such systems that have gained popularity. Agent Tcl [#!Gray96Agent!#] supports a generalized mobility model where migration is allowed at arbitrary points in execution of the mobile code. This provides greater flexibility and perhaps a more natural programming model than we provide. However, this approach suffers from a few shortcomings. First, it requires modification of the core Tcl distribution - something that is difficult to keep up with over the long run. Unrestricted mobility makes support of fault tolerance and reconfiguration harder to achieve. In contrast, our system restricts mobility and other state changes to handler boundaries and treats handlers as atomic units of execution. By providing such a clean execution model, we simplify the system design and implementation while increasing slightly the burden of the developer using our system. Previously, we had developed a system called Sumatra that supports unrestricted mobility for Java applications by modification of the Java Virtual Machine [#!Ranganathan97Network!#] and many of the design decisions in this system are influenced by the experience gained in the Sumatra exercise. TACOMA [#!Johansen95TACOMA!#] is another Tcl-based mobile agent system that adopts a programming model similar to ours. However, there are some basic difference as outlined below.
In this work, we proposed direct communication (reliable message passing) between Mobile Agents. In our system on_stream_append Handlers ( analogous to "Agents" in other systems ) pass one-way messages to each other reliably ( via MStreams ) rather than meeting to exchange messages, using a blackboard or other RPC-like mechanisms. Cabri et. al. [#!Cabri97Coordination!#] argue that this is not such a good idea for several reasons which make sense in the context of free-roaming disconnected agents. Our system is oriented towards building re-configurable distributed applications rather than supporting free-roaming autonomous entities and hence several of their concerns do not apply.
Aglets [#!Lange98Programming!#], Voyager [#!Voyager98ObjectSpace!#], and Mole [#!Strasser96Mole!#] are Java-based systems that follow a programming model similar to ours. However, our system differs from these systems in the following important ways: (1) Our design philosophy is to incorporate reconfiguration into a distributed system building toolkit rather than support disconnected operation as the fundamental design goal, (2) We have incorporated a peer-to-peer reliable, resilient message delivery protocol that none of these other systems offer and (3) We have a means of restricting system re-configuration and extension using policy Handlers that separate global (system-wide), and local concerns.
Dynamic re-configuration of distributed systems has been considered by Hofmeister and Purtilo [#!Hofmeister91Dynamic!#] using a software bus approach. Their system supports dynamic changes to modules, geometry and structure of a distributed system. However, failure processing and asynchronous message delivery during reconfiguration is not considered.