CS:2820 Notes, Lecture 27

From the point of view of the vehicle in our traffic simulation, a trip looks something like this:

vehicle created at source intersection
loop
    pick outgoing road
    enter outgoing road
    reach end of outgoing road
    arrive at intersection
while it is not a sink intersection
vehicle destroyed at sink

Intersection currentIntersection
Road currentRoad = null;
[vehicle created at source intersection, setting currentIntersection]
loop
    [consult road network to pick outgoing road, setting currentRoad]
    [tell currentIntersection that this car is leaving it]
    currentIntersection = null;
    [tell currentRoad that we are entering it, setting travelTime]
    wait for travel time
    [tell currentRoad that we're exiting it, setting currentIntersection]
    [tell currentIntersection that we've arrived, it may make us wait]
    currentRoad = null;
while it is not a sink intersection
vehicle destroyed

The material in square brackets above involves interaction between the code for the vehicle and the code for the road network. We're focusing on the vehicle here, so we'll ignore the details of this interaction.

How can we translate this to an object-oriented discrete-event simulation model? We coluld preserve the above control structure if we were using multithreaded programming, but that is a topic for a class in parallel computing. Here, we'll look at how to do this without going to multiple threads.

Change to an Event Driven Framework

In an event driven framework, each logical process is denied the right to its own control structure. Instead, each segment of the process schedules the next segment. Let's explore this with class Vehicle:

class Vehicle {
    Intersection currentIntersection
    Road currentRoad = null;
    Vehicle( float t, Intersection i ) { // constructor
        currentIntersection = i;
        schedule( t, (float time)->this.pickOutgoing( time ) );
    }

    // head of loop
    void pickOutgoing( float t ) {
        // called when vehicle is in an intersection
        [consult road network to pick outgoing road, setting currentRoad]
        [tell currentIntersection that this car is leaving it]
        currentIntersection = null;
        [tell currentRoad that we are entering it, setting travelTime]
        schedule( t + travelTime, (float time)->this.leaveRoad( time ) );
    }

    // tail of loop
    void leaveRoad( float t ) {
        // called when vehicle arrives at the end of a road
        [tell currentRoad that we're exiting it, setting currentIntersection]
        currentRoad = null;
        [tell currentIntersection that we've arrived, it may make us wait]
        if is not a sink intersection {
            schedule( t, (float time)->this.pickOutgoing( time ) );
        } else {
            vehicle destroyed
        }
    }
}

Other EventDriven Frameworks

Window managers are typically written with mouse events, keypress events, and so on, so applications are written as collections of methods that are called when these events occur. The top level structure of many GUI based applications looks like this:

[ first initialize everything ]
while (true) {
    WindowEvent e = gui.nextEvent();
    switch (e.type) {
	case mouseMove: e.widget().mouseMove( e.mouseCoords() ); break;
	case mouseClick: e.widget().mousePress( e.mouseButton() ); break;
	case mouseClick: e.widget().mouseRelease( e.mouseButton() ); break;
	case keyPress: e.widget().keyPress( e.key() ); break;
	case keyRelease: e.widget().keyRelease( e.key() ); break;
    }
}

In the above, the GUI software uses its model of screen focus to determine which widget on the screen to deliver the event to. Windows, push buttons, text-editing boxes and similar things are all subclasses of widget. The event notice includes what widget the event occurs in, and what happened in that widget, a mouse move, mouse button click, key press, or key release. The main loop calls the appropriate method of the widget for the particular kind of event, and passes that method the details of the event such as where was the mouse, what mouse button was pressed, or what key on the keyboard was involved. It is up to the programmer to write the event service routines in each widget to make the widget behave in a coherent way in response to the sequence of events reported to it.

Transaction processing servers are typically written with an event-driven structure. When you fill out a web form, you are working locally on your client machine until you hit the submit button. At that point, the completed form is delivered, as a single lump, with a "form completed" event. There is typically one event handler for each type of form that a user might fill out.

The notion of a particular client process state is created by variables maintained on the server or, in some cases, by cookies stored on the client's machine and submitted to the server with the form. The logical process, from the client perspective, is the sequence of forms that the user fills out as they visit a the web site managed by that server.

A Problem

Consider the following bit of code from the process-oriented version of the traffic simulator:

class Vehicle {
    ... details omitted ...

    // tail of loop
    void leaveRoad( float t ) {
        // called when vehicle arrives at the end of a road
        [tell currentRoad that we're exiting it, setting currentIntersection]
/**/    [tell currentIntersection we're here, it may make us wait ]
        currentRoad = null;
        if is not a sink intersection {
            schedule( t, (float time)->this.pickOutgoing( time ) );
        } else {
            vehicle destroyed
        }
    }
}

        [tell currentRoad that we're exiting it, setting currentIntersection]

Assuming that we write the appropriate methods in the Road class, we can translate this to Java code something like the following:

        currentIntersection = currentRoad.exit();

The one piece of code above that does not give way to this simple interpretation is the following:

/**/    [tell currentIntersection we're here, it may make us wait ]

The problem with this code is that it involves inter-process interaction. Each intersection can be seen as a process, waiting for one vehicle to clear the intersection before the next enters, or waiting for the stoplight to change, and each vehicle is similarly a process, traveling down roads and through intersections. These processes must interact each time a vehicle reaches an intersection, so this is an example of inter-process interaction.

Inter-Process Interaction

The most fundamental of inter-process interaction involves one processfwaiting for some other process to do something before it continues. Interprocess interactions are common in many contexts: multithreaded programs, interaction between programs running under an operating system, interaction between programs running on a multiprocessor computer or a computer with a multicore processor, and interaction between programs that communicate over a computer network.

For example, in a network, a client program may send a message to a server, requesting a service. Typically, the client then waits for the server to reply. In our case the vehicle is acting in the role of the client, asking the intersection for permission to proceed, but we have no message passing facility.

On a multicore or multiprocessor computer, or in a multithreaded programming environment, one way for one process to wait for another is to use a shared variable:

Process A {
    ...
    v = true;
    while (v) { /* wait */ }
    ...
}

Process B {
    ...
    v = false; /* signal that A can continue */
    ...
}

In general, this is extremely unsafe. Roll-your-own solutions to interprocess interaction are very difficult to debug. This kind of code was typical of how people worked in the 1960s, before general frameworks for interprocess interaction were developed.

Using the Scheduler

In our case, we have a scheduler in our simulation framework, and we can use its interface in a sensible way. When a logical process has to wait, for another process, the code for that process must be broken at the point where the wait occurs so that the second half may be scheduled when it is supposed to occur. Here is what this does to the code we began with:

class Vehicle {
    ... mostly unchanged ...

    // tail of loop, first half
    void leaveRoad( float t ) {
        // called when vehicle arrives at the end of a road
        currentIntersection = currentRoad.exit( t, v );
        currentIntersection.enter( t, currentRoad, /* WHAT GOES HERE */ );
    }

    // tail of loop, second half
    void finishLeaveRoad( float t ) {
        currentRoad = null;
        if is not a sink intersection {
            schedule( t, (float time)->this.pickOutgoing( time ) );
        } else {
            vehicle destroyed
        }
    }
}

The comment /* WHAT GOES HERE */ in the above code is best answered by looking in the code we'll need inside implementations of Intersection:

class SomeKindOfIntersection extends Intersection {
    ...
    void enter( float t, Road r, Something s ) {
        // A vehicle wants to enter this intersection from road r at time t
        if (allowed) {
            // the intersection is clear
            Simulator.schedule( t, s );
        } else {
            // the intersection is blocked
            getqueue( r ).add( s );
        }
    } 
}

Here, we see that, if the intersection is clear, whatever was passed as a parameter is scheduled, while if it is blocked, that same thing is queued up in some queue that depends on the incoming road until such time as the intersection is clear. Clearing the intersection will depend, for example, on some other vehicle leaving the intersection, or on the state of the stoplight changing, or on some similar detail.

So, the thing we pass to the intersection will be the same as the thing we pass to the scheduler! This means the code in class vehicle will look like this:

    // tail of loop, first half
    void leaveRoad( float t ) {
        // called when vehicle arrives at the end of a road
        currentIntersection = currentRoad.exit( t, v );
        currentIntersection.enter( t, currentRoad,
            (float time)->finishLeaveRoad( t )
        );
    }

And, it means that the type of the class of the final parameter to the Intersection.enter method is Simulator.Action. That, is, the correct header for that method is:

class SomeKindOfIntersection extends Intersection {
    ...
    void enter( float t, Road r, Simulator.Action s ) {
        ...
    }
}

It took a long time for people to invent a systematic way of doing this back in the late 1960s. Sadly, the general solution, invented by a student of E. W. Dijkstra in the Netherlands, was published after significant parts of Unix were developed, and as a result, these ideas made it into Unix (and Linux, which is Unix compatible) as afterthoughts. Design by afterthought almost always produces messy results that are never used in many applications that need them.

What we really ought to do is introduce tools into our class Simulation to package the ideas above. The classic packaging, dating back to the time of Dijkstra, is as a new public scheduler class, Semaphore (named after railroad semaphore signals) with two methods, wait() and signal().

If s is a semaphore and a is a semaphore action, then s.wait(a) either schedules a immediately or it sets a aside until it can be done. s.signal() tells the semaphore that if some action is waiting, it can go. What Dijkstra's student discovered is that the semaphore should behave, logically, like an integer. To a user, it is as if the wait operaiton decrements the integer, guaranteeing that it will never go negative by waiting. To a user, it is as if the signal operation increments the integer, which might allow a blocked action to continue.

Much more can be said about semaphores, and if you take an operating systems course or a parallel programming course, you will see more on this.

In fact, in the code we wrote for the road-network simulator, the logic of semaphores is hidden inside class StopLight. We did not call it a semaphore, nor did we make a general mechanism, but we could have done so.

Here is a proposed implementation of this idea as it might be added to the λ expression version of class Simulator:

public abstract class Simulator {
    public interface Action {
	void trigger( double time );
    }
    static void schedule( double time, Action act ) {
	Event e = new Event();
	e.time = time;
	e.act = act;
	eventSet.add( e );
    }
    // other parts of the framework are omitted above

    public class Semaphore {
	private int count;                       // count of excess signals
	private final LinkedList queue;  // queue of excess waits

        public Semaphore( int c ) {
	    count = 1;
	    queue = new LinkedList();
	}

        public void signal( double time ) {
	    if (!queue.isEmpty()) {
	        schedule( time, queue.remove() );
	    } else {
	        count++;
            }
	}

        public void wait( double time, Action act ) {
	    if (count > 0) {
	        count--;
	        schedule( time, act );
	    } else {
	        queue.add( act );
            }
	}
    }

This addition to the simulation framework allows, for example, intersections to have a semaphore used to allow just one vehicle in the intersection at a time. This semaphore would have a count that is initially one. To enter the intersection, a vehicle waits on this semaphore. Once the vehicle is unblocked by this semaphore, it signals when it leaves the intersection. That signal either sets the count back to one or lets the next vehicle in.

Packaging that logic as part of the framework reduces the complexity of intersections significantly.

27. Process Oriented Simulation

The Car's Perspective

Change to an Event Driven Framework

Other EventDriven Frameworks

A Problem

Inter-Process Interaction

Using the Scheduler