3.1. Role of the activity library in Swarm

The activity library provides a foundation for dynamic, object-oriented simulation in Swarm. Swarm assumes that a user defines an object-oriented representation for the structure of a world to be simulated. Once such a representation is established, activity library components are responsible for generating all state changes and flow of information within it. These state changes must be carefully controlled to ensure that they occur only at appropriate times and places in the model. The activity library provides mechanisms to establish this control.

Once the simulation of a dynamic model begins, everything that happens to the model occurs as a direct result of messages sent to world objects by activity library components. These components are also represented by objects, since this is the most effective means for their representation as well, but their purpose is not to represent the current state of the simulated world, but rather to generate changes in world objects by sending messages in a valid order. The activity library has two basic categories of components: those that represent messages to be sent, including constraints on permissible order for sending them, and those that execute the message sends these representations specify, making sure they conform with all constraints.

All changes to a model occur as a result of messages sent to objects within it. These messages invoke compiled methods of receiving objects, which may change local state or send messages to other objects to propagate effects as widely as needed. Because the activity library finally interacts with the world model just by invoking its defined methods, the model can prepackage as much behavior as it can in the form of fast compiled methods. The activity library just triggers the prepackaged behavior at proper times under its own explicit, dynamically interpreted representation.

As conditions shift during execution of a model, the model can also examine and alter the representation of its future behavior contained in the activity structures that drive it. Any changes to this future behavior, however, still occur only as a result of some currently executing action initiated by the activity library.

3.2. Activity library components

The activity library defines one set of basic object types to provide a very rich representation of the kinds of message sending patterns it could generate. Its other major category of object types controls and tracks a current state of processing within these representations.

Both these representations are more abstract than typical object representations, since they deal not with any constant state which can be statically analyzed, but with shifting patterns of messages to be fired to generate such state. While the basic structure of a Swarm simulation is that of a discrete event simulation, its activity representation is also more complicated than most such systems.

There are two basic reasons for the potential complexity of a Swarm activity representation. One is that Swarm supports a decentralized representation of activity to be generated, both to reflect the nature of much of the behavior that it simulates, and so that execution may be distributed across multiple parallel processors. To avoid any need for synchronizing its decoupled activities more often than necessary, Swarm enables a model designer to avoid overspecifying constraints on the patterns of message sends which might potentially be valid. The partial order representation it adopts for a distributed and decoupled plan of activity is inherently more complicated than the single centralized event list adopted by many discrete event simulation systems.

The other reason for potential added complexity is that Swarm's representation of intended activity may be broken into many separate, modular components, which can be bound together in various ways to create larger components. The Swarm composition structure, when fully implemented, will be fundamentally as powerful as the modular abstractions found in most programming languages, with the added complexity of controlling the time at which various events occur.

In spite of this high potential complexity, none of these features is needed for many simple models, such as those that contain only a small variety of basic behaviors, or which define all their behavior to occur at regular, repeating timesteps. The Swarm representation is extremely rich to enable it to scale to large, multi-level models with a variety of dependent behaviors built into the model, and also to facilitate running models on massively parallel machines. Many of the features which seem complex, moreover, are also especially well-suited to building modular and reusable library components. It is anticipated that many of the more advanced features will find their heaviest usage in pre-built libraries that hide internal complexity from applications that use them.

No matter how complex the structures built from them, all the activity library components finally result in the direct execution of messages sends to objects in a model. Because of this direct execution, very precise meaning can be defined for each of them, in terms of message sends that can or must occur. Every component of a structure to be executed, and every event of its execution, can also be accessed using the interfaces of an object-oriented representation. This means that tools can be built to understand and interact with any components to be executed, and also to trace and debug activity as it occurs. None of these tools have been built yet as part of Swarm, but many of the activity library interfaces are present to support them, not because normal use of the library requires the added components.

3.3. Action plan components

The first set of activity library components represent messages to be sent to objects in a model, together with constraints in the order in which they may be sent. All these components are defined as subtypes of the one generic type called ActionPlan.

The two basic kinds of action plans are a simple group of actions to be performed in some order, called an ActionGroup, and a series of actions to be performed at discrete points in time, called a Schedule. The basic element of an action plan is a simple object called an Action. An action defines a particular message to be sent to an object or objects.

In its representation of action plans, Swarm relies heavily on the dynamic message sending capabilities of the Objective C language. The support of Objective C for dynamic message sends is absolutely crucial to Swarm's implementation of generic activity structures. Objective C defines a special kind of data value called a selector, which identifies a message according to its name. One of these selector values is stored in each action of an action plan. During execution of the plan, Objective C machinery performs the actual message send using its selector.

Action plans are free-standing objects that may be created directly by a user program, using a variety of selectable create-time options. Once created, action plans may also refer to each other, by containing special kinds of actions to start another action plan or to perform all the actions within it.

Individual actions can be created only as completely controlled components of some action plan. They are created not by a standard create message, but by special messages (each containing the phrase createAction in its name) sent to an action plan that creates the action as part of the action plan. If the same action is needed in more than one plan, it has to be created in each plan where it is needed. If the entire action plan is dropped, all its actions are also automatically dropped. So the only action plan components which need to be directly managed by an application are the action plans themselves.

Action plans are relatively straightforward components, because they are entirely passive. The only actions they ever contain are those which an application has created in them. These actions stay in them indefinitely unless a special option is requested to clean them out as each one is executed. Because these plans are passive, read-only components to all execution machinery, if the same pattern of actions needs to be triggered at multiple points in a model, it's perfectly valid to create one copy of the actions in an action plan, and then refer to the plan anywhere the actions may be needed.

Even though action plans are passive, containing only what has been placed in them, there is no requirement that their contents remain fixed. Both of the action groups and schedules are implemented directly as collections of their actions. New actions may be added at any time, and existing ones may be dropped or moved from position to another. The contents of action plans may be as dynamic as they need to be to represent the future actions needed in a model. None of these shifting contents has any effect on the model until actually processed during model execution.

3.4. Model execution component

The main responsibility of model execution is just to perform the actions specified by an action plan, in an order of execution consistent with any requirements of the plan. Each action plan may specify as few or as many constraints as it likes over the possible order in which its actions could be performed. Given these specifications, the execution objects are entirely responsible for selecting each action to be performed and then performing them.

There are two simple cases of ordering that account for most all usage, including that of the current Swarm demo programs. One is to permit the actions to be performed in any order, including concurrent execution if hardware and software support were available to do this. The other is to require them to be performed in some specific sequence, one after another, so that the effects of one action could depend on actions which preceded it. This sequence could be either fixed and predetermined, or dynamically established each time the action plan is performed. If a dynamically determined sequence is needed, perhaps even selecting which actions are to be performed at all, users can provide custom subclasses for an action plan and an execution object that performs it. A builtin option is to generate an entirely random sequence each time a plan is executed.

Each time an action plan is being processed, a special kind of object called an activity is created entirely automatically by the runtime processing machinery. These activity objects implement the internal machinery of a virtual processor which has the ability to execute action plans. To get any activity started on a model, a processor is first initialized to run a single top-level plan. All other activity during the lifetime of a model must occur as a result of actions initiated by this plan (which may include the startup of other plans).

Because the activity objects are internal to the processing machinery, an application can usually just ignore their existence. They come and go dynamically as various action plans are started and completed, all arranged in a stack or tree of current activities controlled by a single top-level processor. The activity objects are potentially useful, however, to obtain information about the context in which an action is currently running, or to build debugging or tracing tools to understand actions being performed.

One of the kinds of context information available from an activity object is the current time of a clock value incremented as a schedule is processed. A schedule is a kind of action plan in which all actions occur at specific points in time explicitly established within the schedule. As a schedule is executed, the activity object keeps a current time clock, which holds a global time from the start of all model execution, regardless of the time when the schedule itself was started. An activity object that processes a schedule is called a timeline activity, because its time continually increases from a global base time regardless of times contained in the schedule.

New activities are created whenever an action plan being processed contains an action to perform another action plan, or to start another action plan for autonomous execution. If one action plan performs another, its own processing is stopped until a new activity processing the other action plan completes. If one action plan starts another, the new action plan is started as an autonomous activity controlled only by a higher-level, containing activity.

A Swarm is an activity that exists only to control and coordinate other started subactivities. Unless the subactivities have some special form of explicit synchronization, none of their internal actions has any required ordering relative to those of other plans except as explicitly established during activity execution. The swarm can serve as a simple container of started subactivities which only occasionally synchronize for messages they send to each other. The swarm can be used to hold collections of objects as well as its subactivities as needed to help them coordinate with one another.

One special form of synchronization within a swarm is built into the virtual processing machinery. This synchronization interleaves the actions that occur for every successive time value during processing of timeline subactivities. This merging of actions is often relied on to interleave display and analysis processing with the scheduled actions of a base model. Since no other mechanisms for subactivity coordination are implemented in the current version of Swarm, synchronizing subschedule activities is the major current role of the Swarm activity type. In later versions of Swarm, a swarm will also serve as an important means for organizing a large models into clusters of more densely interacting components, and will also provide a basis of decomposition for parallel execution.