A network model is a set of atomic models and other network models that are interconnected. Network models are used to model large systems that have many interacting parts. Because network models can be components of other network models, it is possible to build models of very large multi-level systems in an organized fashion.
Unlike atomic models, network models do not directly define new dynamic behavior. The dynamics of a network model are determined by the dynamics of its component parts and their interactions. Atomic models define basic dynamic behavior and network models define structure. Separating a model into dynamic behavior and structure greatly aids the task of simulating large systems with many kinds of interacting parts.
The route method is the real workhorse of any Network model. It describes three things. The first is how components of the Network model are connected to each other. The second is how input to the Network model is converted into input for the component models. The third is how output from the component models become output from the Network model.
The signature of the route method is
void route(const X& value, Devs<X>* model, Bag<Event<X> >& r)where the value argument is the event being routed, the model argument is the Network or Atomic model that originated the event, and the r argument is a bag to be filled with the event targets. Each target is described by an Event object that has the target model and the value to be delivered to it. The simulator uses the route method to convert output events produced by Atomic models to, ultimately, input events for other Atomic models. This conversion is done by a somewhat indirect process in which the route method plays a central role.
An example is the easiest way to understand how the simulator uses the route method. The simplest example is converting the output from one Atomic component of the Network into an input for another Atomic component in the same Network. Figure illustrates this case.
The simulator begins by invoking the output_func method of Atomic model
. Next, the simulator iterates through the elements of
's output bag and calls the Network's route method for each one. The arguments passed to route at each call are
void route(const int& value, Devs<int>* model, Bag<Event<int> >& r) { if (model == A) { Event<int> e(B,value); r.insert(e); } }where
It is also possible for the Network model itself to receive input. This can happen when the network is a component in another Network model. Suppose that input to our example Network model becomes input to Atomic model
. Figure
extends Fig.
to include this connection.
When an event appears at the input of the network, the simulator calls the Network's route method with the following arguments:
void route(const int& value, Devs<int>* model, Bag<Event<int> >& r) { if (model == A) { Event<int> e(B,value); r.insert(e); } else if (model == this) { Event<int> e(A,value); r.insert(e); } }
To complete the example, let's extend the network shown in Fig.
void route(const int& value, Devs<int>* model, Bag<Event<int> >& r) { if (model == A) { Event<int> e(B,value); r.insert(e); } else if (model == this) { Event<int> e(A,value); r.insert(e); } else if (model == B) { Event<int> e1(this,value); Event<int> e2(A,value); r.insert(e1); r.insert(e2); } }
The getComponents method is the only other method that must be implemented by a Network subclass. The simulator passes to this method an empty Set of model pointers which must be filled with pointers to the network's components. The getComponents method signature is
void getComponents(Set<Devs<X>*>& c)where c is the set to be filled. There isn't much else to say about this method. The code below shows how it is implemented for the two component network shown in Fig.
void getComponents(Set<Devs<int>*>& c) { c.insert(A); c.insert(B); }
There are just three other items to mention in relation to Network models. First, components should not be connected to themselves. This means that direct feedback loops and direct throughs in a network model must be avoided. These two cases are illustrated in Fig. . Second, direct coupling can only occur between components belonging to the same network, and every component must belong to, at most, one network. Third, you'll notice that it is possible for the route method to modify the value of an output before sending it along. This is permitted and can be useful in some cases.
Conveniently, there are no special rules for simulating networks of network models. The simulator considers the entire collection of atomic models when determining the next event time, output events from atomic models are recursively routed to atomic destinations, and state transitions and garbage collection are performed over the complete set of active atomic components. Hierarchies of network models are a convenient organizing tool for the modeler, but the simulator ultimately treats a multi-level network as a flat structure.
Algorithm is a sketch of the network model simulation procedure. The atomic model simulation algorithm from section
is embedded in the network simulation algorithm. The rules for atomic models do not change in any way; each atomic model sees a sequence of input events and produces a sequence of output events just as before. The only difference here is that the input events are created by other atomic models, and so the input sequence for each atomic model is constructed as the simulation progresses.
I'll use the Adevs SimpleDigraph class to illustrate the construction process. The SimpleDigraph models a network of components whose connections are represented with a directed graph. If, for example, component
is connected to component
, then all output events generated by
become input events to
. The SimpleDigraph has two methods for building a network. The add method takes an Atomic or Network model and adds it to the component set. The couple method accepts a pair of component models and connects the output of the first to the input of the second. Below is the class definition for the model; note that is has a template parameter for setting the input/output type. The Network, Devs, Bag, and Set are in the adevs namespace, and adevs:: must precede them unless the SimpleDigraph is in the adevs namespace (which it is).
template <class VALUE> class SimpleDigraph: public Network<VALUE> { public: /// A component of the SimpleDigraph model typedef Devs<VALUE> Component; /// Construct a network with no components SimpleDigraph():Network<VALUE>(){} /// Add a model to the network. void add(Component* model); /// Couple the source model to the destination model void couple(Component* src, Component* dst); /// Assigns the model component set to c void getComponents(Set<Component*>& c); /// Use the coupling information to route an event void route(const VALUE& x, Component* model, Bag<Event<VALUE> >& r); /// The destructor destroys all of the component models ~SimpleDigraph(); private: // Component model set Set<Component*> models; // Coupling information std::map<Component*,Bag<Component*> > graph; };The SimpleDigraph has two member variables. Pointers to components of the network are stored in the Set models. The components can be Atomic objects, Network objects, or both. The SimpleDigraph components are the nodes of the directed graph. The links, or edges, are stored in the map graph.
The add method does three things. First, it checks that the network is not being added to itself; this is illegal and would cause no end of trouble for the simulator. Next, it adds the component to its component set. Last, the SimpleNetwork sets the component's parent. The last step is needed so that the simulator can climb up and down the model tree. If it is omitted then event routing is likely fail. Here is the implementation of the add method.
template <class VALUE> void SimpleDigraph<VALUE>::add(Component* model) { assert(model != this); models.insert(model); model->setParent(this); }
The couple method does two things, but one of them is somewhat superfluous. First, it adds the source (src) and destination (dst) models to the component set. We could simply have required that the user call the add method before using the couple method, but adding the components here doesn't hurt and might prevent a few headaches. The second step is essential; the method adds the src
dst link to the graph. Notice that the SimpleDigraph itself is a node in the network (but it is not in the component set!). Components that are connected to the network create network outputs. A network connection to a component means that the component will receive network inputs. Here is the couple method implementation.
template <class VALUE> void SimpleDigraph<VALUE>::couple(Component* src, Component* dst) { if (src != this) add(src); if (dst != this) add(dst); graph[src].insert(dst); }
Of the two required methods, route is the more complicated. The arguments to the method are an input event, the network element (i.e., either the SimpleDigraph or one of its components) that is the event source, and the Bag that must be filled with Event objects that indicate the event receivers. The method begins by finding the collection of components that are connected to the event source. Next we iterate through this collection and for each receiver add an Event to the event receiver Bag. When this is done the method returns. The implementation is below.
template <class VALUE> void SimpleDigraph<VALUE>::route(const VALUE& x, Component* model,Bag<Event<VALUE> >& r) { // Find the list of target models and ports typename std::map<Component*,Bag<Component*> >::iterator graph_iter; graph_iter = graph.find(model); // If no target, just return if (graph_iter == graph.end()) return; // Otherwise, add the targets to the event bag Event<VALUE> event; typename Bag<Component*>::iterator node_iter; for (node_iter = (*graph_iter).second.begin(); node_iter != (*graph_iter).second.end(); node_iter++) { event.model = *node_iter; event.value = x; r.insert(event); } }
The second required method, getComponents, is trivial. If we had used some collection other than an Adevs Set to store the components, then the method would have needed to explicitly insert every component model into the Set c. But because models and c are both Set objects, and the Set has an assignment operator, a simple call to that operator is sufficient.
template <class VALUE> void SimpleDigraph<VALUE>::getComponents(Set<Component*>& c) { c = models; }
The constructor and the destructor complete the class. The constructor implementation appears in the class definition; it only calls the superclass constructor. The destructor deletes the component models. Its implementation is shown below.
template <class VALUE> SimpleDigraph<VALUE>::~SimpleDigraph() { typename Set<Component*>::iterator i; for (i = models.begin(); i != models.end(); i++) { delete *i; } }
// Create a digraph model whose components use PortValue<Customer*> // objects as input and output objects. adevs::Digraph<Customer*> store; // Create and add the component models Clerk* clrk = new Clerk(); Generator* genr = new Generator(argv[1]); store.add(clrk); store.add(genr); // Couple the components store.couple(genr,genr->arrive,clrk,clrk->arrive);
The Digraph model is part of the Adevs simulation library. Models that are part of a Digraph must use the adevs::PortValue objects as their input and output type. The Digraph class is a template class with two template parameters. The first is the type of object that will be used as a value in a PortValue object. The second parameter is the type of object that will be used as a port in the PortValue object. The port parameter is of type 'int' by default.
The Digraph class has two primary methods. The add() method is used to add component models to the block diagram model. The couple() method is used to connect components of the Digraph model. The first two arguments to the couple method are the source model and source port. The second two arguments are the destination model and the destination port.
The effect of coupling a source model to a destination model is that
output produced by the source model on the source port appear
as input to the destination model on the destination port. To
illustrate this, consider the output function of the Generator model
shown in Fig. .
void Generator::output_func(Bag<IO_Type>& yb) { // First customer in the list is produced as output IO_Type output(arrive,arrivals.front()); yb.insert(output); }
This places an output value of type ;SPMlt;Customer*> on the ``arrive" output port of the Generator. A corresponding PortValue object appears in the input bag of the Clerk. The value of this PortValue object points to the Customer* object created by the Generator and the port is the Clerk's ``arrive" port.
In addition to coupling Atomic models, the Digraph class can also have other Network models as its components. Suppose that we want to model a convenience store
that has two checkout clerks. When customers are ready to pay their
bill, they look for the line with the smallest number of people and
enter that line. We can reuse the Clerk, Generator, and Observer
models that were introduced in section to build this new model.
The header and source code for the model of the customer's line-selection process is shown below. The model has two output ports, one for each line. There are three input ports. One of these accepts new customers. The others are used to keep track of the number of customers in the each line. The state transition and output functions are self explanatory. Here is the class definition
#include "adevs.h" #include "Customer.h" #include <list> // Number of lines to consider. #define NUM_LINES 2 class Decision: public adevs::Atomic<IO_Type> { public: /// Constructor. Decision(); /// Internal transition function. void delta_int(); /// External transition function. void delta_ext(double e, const adevs::Bag<IO_Type>& x); /// Confluent transition function. void delta_conf(const adevs::Bag<IO_Type>& x); /// Output function. void output_func(adevs::Bag<IO_Type>& y); /// Time advance function. double ta(); /// Output value garbage collection. void gc_output(adevs::Bag<IO_Type>& g); /// Destructor. ~Decision(); /// Input port that receives new customers static const int decide; /// Input ports that receive customers leaving the two lines static const int departures[NUM_LINES]; /// Output ports that produce customers for the two lines static const int arrive[NUM_LINES]; private: /// Lengths of the two lines int line_length[NUM_LINES]; /// List of deciding customers and their decision. std::list<std::pair<int,Customer*> > deciding; /// Delete all waiting customers and clear the list. void clear_deciders(); /// Returns the arrive port associated with the shortest line int find_shortest_line(); };and here is the implementation
#include "Decision.h" #include <iostream> using namespace std; using namespace adevs; // Assign identifiers to ports. Assumes NUM_LINES = 2. // The numbers are selected to allow indexing into the // line length and port number arrays. const int Decision::departures[NUM_LINES] = { 0, 1 }; const int Decision::arrive[NUM_LINES] = { 0, 1 }; // Inport port for arriving customer that need to make a decision const int Decision::decide = NUM_LINES; Decision::Decision(): Atomic<IO_Type>() { // Set the initial line lengths to zero for (int i = 0; i < NUM_LINES; i++) { line_length[i] = 0; } } void Decision::delta_int() { // Move out all of the deciders deciding.clear(); } void Decision::delta_ext(double e, const Bag<IO_Type>& x) { // Assign new arrivals to a line and update the line length Bag<IO_Type>::const_iterator iter = x.begin(); for (; iter != x.end(); iter++) { if ((*iter).port == decide) { int line_choice = find_shortest_line(); Customer* customer = new Customer(*((*iter).value)); pair<int,Customer*> p(line_choice,customer); deciding.push_back(p); line_length[p.first]++; } } // Decrement the length of lines that had customers leave for (int i = 0; i < NUM_LINES; i++) { iter = x.begin(); for (; iter != x.end(); iter++) { if ((*iter).port < NUM_LINES) { line_length[(*iter).port]--; } } } } void Decision::delta_conf(const Bag<IO_Type>& x) { delta_int(); delta_ext(0.0,x); } double Decision::ta() { // If there are customers getting into line, then produce output // immediately. if (!deciding.empty()) { return 0.0; } // Otherwise, wait for another customer else { return DBL_MAX; } } void Decision::output_func(Bag<IO_Type>& y) { // Send all customers to their lines list<pair<int,Customer*> >::iterator i = deciding.begin(); for (; i != deciding.end(); i++) { IO_Type event((*i).first,(*i).second); y.insert(event); } } void Decision::gc_output(Bag<IO_Type>& g) { Bag<IO_Type>::iterator iter = g.begin(); for (; iter != g.end(); iter++) { delete (*iter).value; } } Decision::~Decision() { clear_deciders(); } void Decision::clear_deciders() { list<pair<int,Customer*> >::iterator i = deciding.begin(); for (; i != deciding.end(); i++) { delete (*i).second; } deciding.clear(); } int Decision::find_shortest_line() { int shortest = 0; for (int i = 0; i < NUM_LINES; i++) { if (line_length[shortest] > line_length[i]) { shortest = i; } } return shortest; }
The block diagram model of the store with multiple clerks is shown in Fig. . The
external interface for this block diagram model is identical to the
previous clerk models (i.e., the Clerk and Clerk2 models), and we can
use the generator and observer models to conduct the same experiments
as before. The external ``arrive" input of the
multi-clerk model is connected to the ``decide" input of
the Decision model. The ``depart" output ports of each of
the Clerk models is connected to the external ``arrive"
output port of the multi-clerk model. The Decision model has two
output ports, each one producing customers for a distinct clerk. These
output ports are coupled to the ``arrive" port of the
appropriate clerk model. The Clerk's ``depart" output ports are
then coupled to the appropriate ``departure" port of the
decision model.
The multi-clerk model is implemented by deriving a new class from the Digraph class. The constructor of the new class creates and adds the component models and establishes their interconnections. Here is the header file for the new multi-clerk model.
#include "adevs.h" #include "Clerk.h" #include "Decision.h" /** A model of a store with multiple clerks and a "shortest line" decision process for customers. */ class MultiClerk: public adevs::Digraph<Customer*> { public: // Model input port static const int arrive; // Model output port static const int depart; // Constructor. MultiClerk(); // Destructor. ~MultiClerk(); };And here is the source file
#include "MultiClerk.h" using namespace std; using namespace adevs; // Assign identifiers to I/O ports const int MultiClerk::arrive = 0; const int MultiClerk::depart = 1; MultiClerk::MultiClerk(): Digraph<Customer*>() { // Create and add component models Decision* d = new Decision(); add(d); Clerk* c[NUM_LINES]; for (int i = 0; i < NUM_LINES; i++) { c[i] = new Clerk(); add(c[i]); } // Create model connections couple(this,this->arrive,d,d->decide); for (int i = 0; i < NUM_LINES; i++) { couple(d,d->arrive[i],c[i],c[i]->arrive); couple(c[i],c[i]->depart,d,d->departures[i]); couple(c[i],c[i]->depart,this,this->depart); } } MultiClerk::~MultiClerk() { }Notice that the MultiClerk destructor does not delete the component models. This is because the component models are adopted by the base class when they are added to the Digraph. Consequently, the component models are deleted by the base class destructor, rather than the destructor of the derived class.
Our implementation of the Game of Life has two parts: the atomic models that implement the individual cells and the CellSpace model that contains the cells and routes their output events. The CellSpace is a type of Network. The components of a CellSpace exchange CellEvent objects that have four fields: the x, y, and z coordinates of the target cell and a value to deliver. The CellEvent class is a template class whose template argument sets the value type. The size of the CellSpace is determined when the CellSpace object is created, and it has methods for adding and retrieving cells by their location.
The Atomic components in our Game of Life implementation have two state variables: the dead or alive status of the cell and the number of living neighbors. Two methods are implemented to test the death and rebirth rules, and the cell sets its time advance to 1 whenever a rule is satisfied. The cell output is its new dead or alive state. External events update the cell's living neighbor count. In order to produce properly targeted CellEvents, each cell also keeps track of its own location in the cell space. In the example code, the cell space is rendered graphically using OpenGL, but I'll omit that part. Here is the header file for our Game of Life cell.
/// Possible cell phases typedef enum { Dead, Alive } Phase; /// IO type for a cell typedef adevs::CellEvent<Phase> CellEvent; /// A cell in the Game of Life. class Cell: public adevs::Atomic<CellEvent> { public: /** Create a cell and set the initial state. The width and height fields are used to determine if a cell is an edge cell. The last phase pointer is used to visualize the cell space. */ Cell(long int x, long int y, long int width, long int height, Phase phase, short int nalive, Phase* vis_phase = NULL); ... Required Adevs methods and destructor ... private: // location of the cell in the 2D space long int x, y; // dimensions of the 2D space static long int w, h; // Current cell phase Phase phase; // number of living neighbors. short int nalive; // Output variable for visualization Phase* vis_phase; // Returns true if the cell will be born bool check_born_rule() const { return (phase == Dead && nalive == 3); } // Return true if the cell will die bool check_death_rule() const { return (phase == Alive && (nalive < 2 || nalive > 3)); } };
The template argument supplied to the base Atomic class is a CellEvent whose value field has the type Phase. The check_born_rule method tests the rebirth condition and check_death_rule method tests the death condition. The appropriate rule, as determined by the cell's dead or alive status, is used in the time advance, output, and internal transition methods. The number of living cells is updated by the cell's delta_ext method whenever neighboring cells report a change in their health. Here are the Cell's method implementations.
Cell::Cell(long int x, long int y, long int w, long int h, Phase phase, short int nalive, Phase* vis_phase): adevs::Atomic<CellEvent>(),x(x),y(y),phase(phase),nalive(nalive),vis_phase(vis_phase) { // Set the global cellspace dimensions Cell::w = w; Cell::h = h; // Set the initial visualization value if (vis_phase != NULL) *vis_phase = phase; } double Cell::ta() { // If a phase change should occur then change state if (check_death_rule() || check_born_rule()) return 1.0; // Otherwise, do nothing return DBL_MAX; } void Cell::delta_int() { // Change the cell state if necessary if (check_death_rule()) phase = Dead; else if (check_born_rule()) phase = Alive; } void Cell::delta_ext(double e, const adevs::Bag<CellEvent>& xb) { // Update the living neighbor count adevs::Bag<CellEvent>::const_iterator iter; for (iter = xb.begin(); iter != xb.end(); iter++) { if ((*iter).value == Dead) nalive--; else nalive++; } } void Cell::delta_conf(const adevs::Bag<CellEvent>& xb) { delta_int(); delta_ext(0.0,xb); } void Cell::output_func(adevs::Bag<CellEvent>& yb) { CellEvent e; // Assume we are dying e.value = Dead; // Check in case this in not true if (check_born_rule()) e.value = Alive; // Set the visualization value if (vis_phase != NULL) *vis_phase = e.value; // Generate an event for each neighbor for (long int dx = -1; dx <= 1; dx++) { for (long int dy = -1; dy <= 1; dy++) { e.x = (x+dx)%w; e.y = (y+dy)%h; if (e.x < 0) e.x = w-1; if (e.y < 0) e.y = h-1; // Don't send to self if (e.x != x || e.y != y) yb.insert(e); } } }The output_func method shows how a cell sends messages to its neighbors. The double for loop creates a CellEvent targeted at each adjacent cell. The location of the target cell is written to the x, y, and z fields of the CellEvent object. Just like arrays, the location values can range from zero to the cell space size minus one. The CellSpace will do the actual routing of the CellEvents to their targets. Note however that if the target of the CellEvent is outside of the cell space, then the CellSpace itself will produce the CellEvent as an output.
The remainder of the simulation program looks very much like the other simulation programs that we've seen so far (except for some OpenGL specific code, omitted here, that is used to display the cell space). A CellSpace object is created and we add each cell to it. Then a Simulator object is create and a pointer to the CellSpace is passed to the Simulator's constructor. Last, we execute events until our stopping criteria is met. The execution part is already familiar, so let's just focus on creating the CellSpace. Here is the code snippet that performs the construction.
// Create the cellspace model cell_space = new adevs::CellSpace<Phase>(WIDTH,HEIGHT); for (int x = 0; x < WIDTH; x++) { for (int y = 0; y < HEIGHT; y++) { // Count the living neighbors short int nalive = count_living_cells(x,y); // The 2D phase array contains the initial Dead/Alive state of each cell cell_space->add( new Cell(x,y,WIDTH,HEIGHT,phase[x][y],nalive,&(phase[x][y])),x,y); } }Just as with the Digraph class, the CellSpace template argument determines the value type for the CellEvents that are used as input and output by the CellSpace components. The CellSpace constructor sets the dimensions of the space. Every CellSpace is three dimensional, and the constructor accepts three arguments that set the x, y, and z dimensions; omitted arguments default to 1. The constructor signature is
CellSpace(long int width, long int height = 1, long int depth = 1)
Components are added to the cellspace with the add method. This method places a component at a specific (x,y,z) location. Its signature is
void add(Cell* model, long int x, long int y = 0, long int z = 0)where Cell is a Devs (atomic or network) by the type definition
typedef Devs<CellEvent<X> > Cell;The CellSpace deletes its components when it is deleted. The CellSpace class has five other methods for retrieving cells and getting the dimensionality of the cell space. These are more or less self-explanatory; the signatures are shown below.
const Cell* getModel(long int x, long int y = 0, long int z = 0) const; Cell* getModel(long int x, long int y = 0, long int z = 0); long int getWidth() const; long int getHeight() const; long int getDepth() const;
The Game of Life produces a surprising number of clearly recognizable patterns. Some of these patterns are fixed and unchanging; others oscillate, cycling through a set of patterns that always repeats itself; others seem to crawl or fly. One familiar static pattern is the Block shown in Fig. .
Our discrete event implementation of the Game of Life doesn't do any work when simulating a Block. None of the cells in a Block change in any way; their phases are constant and so are their neighbor counts.
The confluent transition function plays a major role in the Blinker simulation. Most of the rows in Table (all but the first row, in fact) have both an input and an output, which means that an internal and external event coincide and so the next state is determined by the delta_conf method. It is also important that the input and output bags carry multiple values. The external transition function (which is used in defining the confluent transition function) must be able to compute the number of living neighbors before determining its next state. If input events were provided one at a time (e.g., if the input bag were replaced by a single input event), then our discrete event Game of Life would be much more difficult to implement.