Proceedings of the 1997 Winter Simulation Conference

ed. S. Andradttir, K. J. Healy, D. H. Withers, and B. L. Nelson

INSIDE DISCRETE-EVENT SIMULATION SOFTWARE:

HOW IT WORKS AND WHY IT MATTERS

Thomas J. Schriber

Daniel T. Brunner

Computer and Information Systems

The University of Michigan
Ann Arbor, Michigan 48109-1234, U.S.A.

Systemflow Simulations, Inc.

6366 Guilford Avenue, Suite 300
Indianapolis, Indiana 46220-1750, U.S.A.

SIMAN, ProModel, and GPSS/H. An expanded version

of the 1996 material containing figures, flow charts, and
additional text is in Schriber and Brunner (1998).

ABSTRACT
This paper provides simulation practitioners and
consumers with a grounding in how discrete-event
simulation software works. Topics include discreteevent systems; entities, resources, control elements and
operations; simulation runs; entity states; entity lists; and
entity-list management. The implementation of these
generic ideas in AutoMod and SLX is described. The
paper concludes with several examples of why it
matters for modelers to know how their simulation
software works, including coverage of SIMAN,
ProModel and GPSS/H as well as the other two tools.

1.2 Structure of the Paper

In Sections 2, 3 and 4 we comment on the nature of
discrete-event simulation; basic simulation consstructs
such as entities, resources, control elements, and operations; and model execution. Sections 5 and 6 deal in a
general way with entity states and entity management
data structures. Section 7 discusses three specific
implementations of entity management rules. Section 8
explores why it matters.

1 INTRODUCTION
1.3 Terminology and Conventions
1.1 Background
Throughout this paper we use terms that we define as
well as terms reserved by the developers of particular
simulation tools. Terms we define are boldfaced on first
use. Tool-specific terms are Capitalized or, where
appropriate, are spelled out in ALL CAPS.

A black box approach is often taken in teaching and

learning discrete-event simulation software. The external
characteristics of the software are studied, but the
foundation on which the software is based is ignored or
is touched on only briefly. Choices made in implementation of the foundation might not be studied at all
and related to step-by-step model execution. The
modeler therefore might not be able to think things
through when faced with such needs as developing good
approaches for modeling complex situations, using
interactive tools to come to a rapid understanding of
error conditions arising during model development, and
using interactive tools to verify that complex system
logic has been captured correctly in a model. The
objective of this paper, then, is to describe the logical
underpinnings of discrete-event simulation and illustrate
this material in terms of various implementations of
discrete-event simulation software.
This paper is a revised version of an identically named
paper from the 1996 Winter Simulation Conference
(Schriber and Brunner 1996). The 1996 paper covered
the entity-list management rules and why it matters for

2 ABOUT DISCRETE-EVENT SIMULATION

2.1 The Transaction-Flow World View
The transaction-flow world view often provides the
basis for discrete-event simulation. In the transactionflow world view, a system is visualized as consisting of
discrete units of traffic that move (flow) from point to
point in the system while competing with each other for
the use of scarce resources. The units of traffic are
sometimes called transactions, giving rise to the phrase
transaction flow.
Numerous systems fit the preceding description.
Included are many manufacturing, material handling,
transportation, health care, civil, natural resource,
communication, defense, and information processing
systems, and queuing systems in general.

14

Inside Discrete-Event Simulation Software: How it Works and Why it Matters

15

A discrete-event simulation is one in which the state of a

model changes at only a discrete, but possibly random,
set of simulated time points. Two or more traffic units
often have to be manipulated at one and the same time
point. Such simultaneous movement of traffic at a
time point is achieved by manipulating units of traffic
serially at that time point. This often leads to logical
complexities in discrete-event simulation because it
raises questions about th e order in which two or more
units of traffic are to be manipulated at one time point.

The term control element designates a construct that

supports other delays or logical alternatives based on a
systems state. Control elements can take the form of
switches, counters, user data values, and system data
values built into the modeling tool. Complex control
may rely on truth-valued expressions that use arithmetic
and/or Boolean combinations of control elements.
An operation is a step carried out by or on an entity
while it moves through a system. The operations applicable to a ship at a harbor might be these: arrive; capture
a berth; capture a tugboat; get pulled into the berth; free
the tugboat; load cargo; etc.

2.3 Discrete-Event Modeling Languages

4 OVERVIEW OF MODEL EXECUTION

The challenges faced by a modeler escalate for the

designer of a modeling language. The designer must
take the logical requirements of discrete-event simulation
into account in a generalize d way. Choices and tradeoffs
exist. As a result, although discrete-event simulation
languages are similar in broad terms, they can and
typically do differ in subtle but important particulars.

4.1 Experiments, Replications, and Runs

2.2 The Nature of Discrete-Event Simulation

3 ENTITIES, RESOURCES, CONTROL

ELEMENTS, AND OPERATIONS
The term entity is used here to designate a unit of traffic
(a transaction). Entities instigate an
d respond to
events . An event is a happening that changes the state of
a model (or system). In a model of an order-filling
system, for example, the arrival of an order, which is an
event, might be simulated by bringing an entity into the
model.
There are two possible types of entities, here referred
to as external entities and internal entities . External
entities are those whose creation and movement is
explicitly arranged for by the modeler. In contrast,
internal entities are created and manipulated implicitly by
the simulation software itself. For example, internal
entities might be used in some languages to simulate
machine failures, whereas external entities might be used
to simulate the use of machines.
The term resource designates a system element that
provides service (such as a drill, an automated guided
vehicle, or space in an input buffer). The users of
resources are usually entities. (A work-in-process entity
claims space in an input buffer, then captures an
automated guided vehicle to move it to the input buffer.)
Resources are usually capacity-limited , so entities
compete for their use and sometimes must wait to use
them, experiencing delay as a result.

A simulation project is composed of experiments .

Experiments are differentiated by the use of alternatives
in a models logic and/or data. An alternate part
sequencing rule might be tried, for example, or the
quantity of various machines might be varied.
Each experiment consists of one or more replications
(trials). A replication is a simulation that uses the experiments model logic and data but a different set of random numbers, and so produces different statistical results
that can then be analyzed across a set of replications.
A replication involves initializing the model, running it
until a run-ending condition is met, and reporting results.
This running it phase is called
a run .
4.2 Inside a Run
During a run the simulatio n clock (an internally managed, stored data value) tracks the passage o fsimulated
time (as distinct from wall-clock time). The clock advances in discrete steps (typically of unequal size) during
the run. After all possible actions have been taken at a
given simulated time, the clock is advanced to the time of
the next earlies t event. Then the appropriate actions are
carried out at this new simulated time, etc.
The execution of a run thus takes the form of a twophase loop: carry out all possible actions at the current
simulated time, followed by advance the simulated
clock, repeated over and over again until a run-ending
condition comes about. The two phases are here
respectively called the Entity Movement Phase (EMP)
and the Clock Update Phase (CUP).

16

Schriber and Brunner

5 ENTITY STATES

6 ENTITY MANAGEMENT STRUCTURES

Entities migrate from state to state while they work their

way through a model. An entity is always in one of five
alternative states, as detailed below.

Simulation software uses the following lists to organize

and track entities in the five entity states.
6.1 The Active Entity

5.1 The Active State

The Active State is the state of the currently moving entity. Only one entity moves at any instant of wall-clock
time. This entity progresses through its operations
nonstop until it encounters a delay. It then migrates to an
alternative state. Some other entity then becomes the
next active entity. And so on.

The active entity forms an unnamed list consisting only

of the active entity. The Active-State entity moves nonstop until encountering an operation that puts it into another state (transfers it to another list) or removes it from
the model. A Ready-State entity then becomes the next
Active-State entity. Eventually there is no possibility of
further action at the current time. The EMP then ends
and a Clock Update Phase begins.

5.2 The Ready State

6.2 The Current Events List
During an Entity Movement Phase there may be more
than one entity ready to move, and yet entities can only
move (be in the Active State) one-by-one. The Ready
State is the state of entities waiting to enter the Active
State during the current Entity Movement Phase.
5.3 The Time-Delayed State
The Time-Delayed State is the state of entities waiting
for a known future simulated time to be reached so that
they can then (re)enter the Ready State. A part entity
is in a Time-Delayed State, for example, while waiting
for the future simulated time at which an operation being
performed on it by a machine will come to an end.
5.4 The Condition-Delayed State
The Condition-Delayed State is the state of entities delayed until some specified condition comes about, e.g., a
part entity might wait in the Condition-Delayed State
until its turn comes to use a machine. Condition-Delayed
entities are removed automatically from the ConditionDelayed state when conditions permit.
5.5 The Dormant State
Sometimes it is desirable to put entities into a state from
which no escape will be triggered automatically by
changes in model conditions. We call this state the
Dormant State. Dormant-State entities rely on modelersupplied logic to transfer them from the Dormant State
back to the Ready State. Job-ticket entities might be put
into a Dormant State, for example, until an operator
entity decides which job-ticket to pull next.

Entities in the Ready State are kept in a single list here

called the current events list (CEL). Entities migrate to
the current events list from the future events list, from
delay lists, and from user-managed lists. (Each of these
latter lists is described below). In addition, entities
cloned from the Active-State entity usually start their
existence on the current events list.
6.3 The Future Events List
Entities in the Time-Delayed State belong to a single list
into which they are inserted at the beginning of their
time-based delay. This list, called the future events list
(FEL) here, is usually ranked by increasing entity move
time. (Move time is the simulated time at which an
entity is scheduled to try to move again.) At the time of
entity insertion into the FEL, the entitys move time is
calculated by adding the value of the simulation clock to
the known (sampled) duration of the time-based delay.
After an Entity Movement Phase is over, the Clock
Update Phase sets the clocks value to the move time of
the FELs highest ranked (smallest move time) entity.
This entity is then transferred from the FEL to the CEL,
migrating from the Time-Delayed State to the Ready
State and setting the stage for the next EMP to begin.
The preceding statement assumes there are not other
entities on the FEL whose move time matches the clocks
updated value. In the case of move-time ties, some tools
will transfer all the time-tied entities from the FEL to the
CEL during a single CUP, whereas other tools take a
one entity transfer per CUP approach.
Languages that work with internal entities usually use
the FEL to support the timing requirements of these
entities. The FEL is typically composed both of external
and internal entities in such languages.

Inside Discrete-Event Simulation Software: How it Works and Why it Matters

17

6.4 Delay Lists

7.1 AutoMod

Delay lists are lists of entities in the Condition-Delayed

State. These entities are waiting for a condition to come
about (e.g., waiting their turn to use a machine) so they
can be transferred automatically into the Ready State on
the curren t events list. Delay lists, which are generally
created automatically by the simulation software, are
managed by using related waiting or polled waiting .
If a delay can be related easily to events in the model
that might resolve the condition, then related waiting can
be used to manage the delay list. For example, suppose a
machines status changes from busy to idle. In response,
the software can automatically remove the next machineusing entity from the appropriate delay list and put it in
the Ready State on the current events list. Related
waiting is the prevalent approach used to manage
conditional delays.
If the delay condition is too complex to be related easily to events that might resolve it, polled waiting can be
used. With polled waiting the software checks routinely
to see if entities can be transferred from one or more
delay lists to the Ready State. Complex delay conditions
for which polled waiting can be useful include Boolean
combinations of state changes, e.g., a part supply runs
low or an output bin needs to be emptied.

AutoMod equivalents for the preceding generic terms are

given in Table 1. For example, AutoMod uses Actions to
specify operations for Loads.

6.5 User-Managed Lists

User-managed lists are lists of entities in the Dormant
State. The modeler must take steps to establish such lists
and provide the logic needed to transfer entities to and
from the lists. (The underlying software has no way to
know why entities are put into user-managed lists and so
has no basis for removing entities from such lists.)
7 IMPLEMENTATION IN TWO TOOLS
The tools chosen for commentary on implementation
particulars
are
AutoMod,
Version
8,
from
AutoSimulations, Inc., and SLX, Release 1, from
Wolverine Software Corporation. (See the References.)
A previous version of this paper (Schriber and Brunner
1996) covered Systems Modeling Corporations SIMAN,
ProModel Corporations ProModel, and Wolverine
Software Corporations GPSS/H in similar detail. These
five are among more than fifty tools reported in 1995 for
discrete-event simulation (Swain 1995). Some other
tools might be better suited than any of these for
particular modeling activities, but we think that these
tools are representative.
(Those interested in the
possibility of implementing discrete-event simulation
models in a non-simulation programming language such
as C or C++ are referred to Balci 1988.)

Table 1: AutoMod Terminology

Generic Term
External Entity
Internal Entity
Resource
Control Element
Operation
Current Events List
Future Events List
Delay List

User-Managed List

AutoMod Equivalent
Load
Logical Load
Resource; Queue; Block
Counter;
Process Traffic Limit
Action
Current Event List
Future Event List
Delay List;
Condition Delay List;
Load Ready List
Order List

7.1.1 The Current Event List

The current events list is named the Current Event List in
AutoMod. Cloned Loads, Loads leaving the Future
Event List due to a clock update, and Loads ordered off
Order Lists are placed immediately on the CEL. The
insertion rule is to rank by priority and then FIFO within
the priority class.
When the CEL becomes empty, the Condition Delay
List (see below) is checked, and Loads may be
transferred from there to the CEL. This continues until
the CEL is empty and no more Loads can be transferred,
at which point the EMP is over and a CUP is initiated.
7.1.2 The Future Event List
The AutoMod Future Event List (FEL) is like future
events lists in other tools. Loads arriv e on the FEL in the
Time-Delayed State by executing a WAIT FOR
statement. AutoMod allows the specification of time
units (day, hr, min, sec) in a WAIT FOR statement.
The AutoMod CUP will remove multiple Transactions
from the FEL if they are tied for the earliest move time,
inserting them one by one into their appropriate place on
the CEL.
There are also internal entities in AutoMod, called
Logical Loads, that do things such as wait on the FEL to
trigger scheduled shift breaks.

18

Schriber and Brunner

7.1.3 Delay Lists

7.1.5 Order Lists

Delay Lists (DL) are lists of loads waiting to claim

capacity of a finite capacity element (a resource or
control element such as an individual Resource, Queue,
Block, Counter, or Process). Each finite capacity
element within the model has one DL associated with it.
The waiting that results from this mechanism is related
waiting. Whenever capacity is freed, one Load from the
head of the elements DL gets tentatively placed on the
CEL (but a placeholder is left on the DL). When that
Load is encountered during the EMP, it tries to claim the
requested capacity. If it fails (for example because it
wants two units but only one is free), it is returned to the
DL in its original place.
Immediately after this evaluation (and before the
active Load executes any more Actions), if there is still
any capacity, the next Load on the DL is placed on the
CEL. Processing of the active Load then continues.
After each time a tentatively placed Load is evaluated
during the EMP, the existence of available capacity will
cause another Load to be removed from the DL.

AutoMod implements the Dormant State with Order

Lists, which are user-managed lists of Loads. After a
Load puts itself onto an Order List (by executing a
WAIT TO BE ORDERED Action), it can only be
removed by another Load (which executes an ORDER
Action).
Loads successfully ordered are placed
immediately on the CEL (one at a time according to how
they were chosen from the Order List, and ranked on the
CEL FIFO by priority).
Order Lists can achieve performance improvements
over CDL waiting because Order Lists are never scanned
except on explicit request.
AutoMod Order Lists offer several interesting wrinkles
including the ability for an ordering Load to place a back
order if the ORDER action is not satisfied, the ability for
a Load on an Order List to be ordered to continue (to the
next Action) instead of to a Process (this feature is useful
for control handshaking), and the ability to have a
function called for each Load on the Order List (by using
the ORDERSATISFYING Action).

7.1.4 The Condition Delay List

7.1.6 Other Lists

For conditional waiting other than the five cases

enumerated above, AutoMod has a WAIT UNTIL
statement that results in polled waiting. WAIT UNTIL
conditions can be compounded using Boolean operators.
If a Load executes a WAIT UNTIL and the condition is
false, the Load is placed on a single global AutoMod list
called the Condition Delay List (CDL).
After the CEL has been emptied, but before the
simulation clock progresses, all loads on the CDL are
moved to the CEL if there has been a state change to at
least one element of the same general type (e.g. Queue)
that any Load on the CDL is waiting for. (This
mechanism is primarily polled although the global
triggering mechanism is related.)
If the CEL is now non-empty then the EMP resumes.
If the condition that a CEL Load is waiting for is false,
AutoMod moves that Load from the Current Event List
back to the CDL. The CDL may get emptied multiple
times during one EMP until eventually the CEL gets
emptied without having triggered a state change related
to any Load on the CDL. A CUP then occurs.
Because of the potential for repetitive list migration
when using WAIT UNTIL, AutoMods vendor
encourages users to use Order Lists or other explicit
control mechanisms to manage complex waiting.

AutoMod has a number of material handling constructs

that are integrated with Load movement. For vehicle
systems there are three other types of lists. Loads on
Load Ready Lists (LRL) (one list per vehicle system) are
waiting to be picked up by a vehicle. Loads claimed and
picked up by a vehicle reside on the vehicles Vehicle
Claim List (VCL) and Vehicle Onboard List (VOL)
respectively, during which time, the vehicle becomes the
active load and moves among AutoMods lists (FEL,
CEL, and possibly DLs) instead of the Load.
7.2 SLX
SLX is a hierarchical language in which the built-in
primitives are at a lower level than most simulation
languages, facilitating user (or developer) definition of
the behavior of many system elements. This philosophy
allows the SLX developer to create higher-level
modeling tools whose constructs have precisely defined
yet modifiable behavior.
Equivalents for the generic terms for users of low-level
SLX are given in Table 2. For example, SLX uses
Control Variables to act as Control Elements. The
control modifier can be attached to a global or local
Variable of any data type (integer, real, string, etc.). (A
local Variable is analogous to an attribute in other tools.)
Note that SLX has two types of Objects: Active and
Passive. An Active Object is distinguished from a

Inside Discrete-Event Simulation Software: How it Works and Why it Matters

Passive Object by the presence of actions executable

Statements in an Active Objects Class definition.
Table 2: SLX Terminology (low-level)
Generic Term
External Entity
Internal Entity
Resource
Control Element
Operation
Current Events List
Future Events List
Delay List
User-Managed List

SLX Equivalent
Active Object and its Puck(s)
none
Control Variable
Control Variable
Statement
Current Events Chain
Future Events List
Delay List
Set (see section 7.2.4)

Table 3 shows how higher-level tools based on SLX

might exploit the definitional capabilities of SLX.
Table 3: Tools Based On SLX
Generic Term
Resource
Control Element
Operation
Delay List
User-Managed List

SLX Equivalent
Active or Passive Object
Active or Passive Object
User-defined Statement
User-defined based on Set
User-defined based on Set

7.2.1 The Current Events Chain

The current events list is named the Current Events
Chain (CEC) in SLX. The members of the CEC are
called Pucks.
What is a Puck? SLX dissociates the concept of an
Active Object (with its associated local data) from a
Puck, which is the moving entity that executes the
actions, carries its own entity scheduling data, and
migrates from list to list. The effect of this dissociation
is that a single Object can own more than one Puck.
All Pucks owned by a single Object share the Objects
local data (attributes). On e applicatio n of this local
parallelism feature is the use of a second Puck to
simulate a balk time while the original Puck is waiting
for some condition (Henriksen 1996).
Activating a new Object creates one Puck and launches
that Puck into the actions. In many cases no additional
Pucks are ever created, and the combination of an Active
Object and its Puck forms the equivalent to an entity in
the terminology of this paper. (Passive Objects have no
actions and therefore own no Pucks.)
Newly activated Pucks, Pucks leaving the Future Event
List due to a clock update, and reactivated Pucks (see
7.2.4 below) are placed immediately into the CEC. The

19

CEC is ranked FIFO by priority. The SLX CEC is empty

when an EMP ends.
7.2.2 The Future Events List
The SLX Future Events List (FEL) is like future events
lists in other tools. Loads arrive on the FEL in the TimeDelayed State by executing an ADVANCE statement.
The SLX CUP will remove multiple Pucks from the
FEC if they are tied for the earliest move time, inserting
them one by one into their appropriate place on the CEC.
Because the low-level primitives in SLX do not
include downtimes or even repetitive Puck generation, all
activity on the SLX FEL unfolds as specified by the
developer of the SLX model. However, if a user is using
a model (or a model builder) that contains higher-level
primitives defined by a developer, chances are that all
kinds of things are going on behind the scenes, hidden
from the higher-level users view.
7.2.3 Delay Lists
Delay Lists (DL) are lists of Pucks waiting (through
WAIT UNTIL) for state changes in any combination of
Control Variables and the simulation clock value. All
higher-level constructs defined by developers can use
this mechanism. Each Control Variable (which may be a
local Variable, in which case there is one fo r each Object
in the Class) has a separate DL associated with it.
A DL is ranked by order of insertion. The entire
contents of a DL are removed whenever the associated
Control Variable changes value and are inserted one at a
time into the CEC. Removed Pucks that ar e waiting on
compound conditions are removed from each other
Delay List to which they belong. As these entities are
encountered on the CEC during the EMP, those failing to
pass their WAIT UNTIL are returned to the Delay List(s)
for those Control Variables still contributing to the
falseness of the condition.
For conditions that include a reference to the clock, the
Puck is inserted if necessary into the FEL, subject to
early removal from the FEL if the condition becomes
true do to other Control Variable changes.
This low-level related waiting mechanism based on
Control Variables is the default SLX approach to
modeling all types of simple or compound ConditionDelayed states.
7.2.4 Sets and User-Managed Waiting
SLX handles the Dormant State in a unique way. Instead
of moving the Puck from the active state to a usermanaged list and suspending it, all in the same operation,
SLX breaks this operation into two pieces.

20

Schriber and Brunner

First, the Puck should join a Set. But joining a Set

does not automatically suspend the Puck. A Puck can
belong to any number of Sets. Set membership merely
provides other Pucks with access to the member Puck.
To go into the Dormant state, a Puck executes a WAIT
statement. It then is suspended indefinitely, outside of
any particular list, until another Puck identifies the
waiting Puck and executes a REACTIVATE statement
on it. Often the REACTIVATEing Puck is scanning a
Set to find the Puck to REACTIVATE, but a Set is not
exactly the same as a user-managed list in our
terminology. A Dormant-state Puck might be a member
of no Sets (as long as a pointer to it has been stashed
somewhere) or of one or more Sets.
An SLX developer can easily define a user-managed
list construct, using Sets, WAIT, and REACTIVATE as
building blocks, that mimics those of other languages or
offer unique features of its own.
8 WHY IT MATTERS
8.1 Overview
We now describe five situations that reveal some of the
practical differences in implementation particulars
among SIMAN, ProModel, GPSS/H, AutoMod, and
SLX. These differences reflect differing implementation
choices made by the software designers.
None of the alternative approaches mentioned in each
subsection is either intrinsically right or wrong. The
modeler simply must be aware of the alternative in effect
in the simulation software being used and work with it to
produce the desired outcome. (If a modeler is unaware
of the alternative in effect, it is possible to mis-model a
situation and perhaps not become aware of this fact.)
We finish the why it matters discussion with some
comments on how knowledge of software internals is
needed to make effective use of software checkout tools.
8.2 Trying to Re-capture a Resource Immediately
Suppose a part releases a machine, then immediately
attempts to re-capture the machine. The modeler might
or might not want a more highly qualified waiting part,
if any, be the next to capture the machine.
Of interest here is the order of events following the
giving up of a server. There are at least three
alternatives: (1) Coupled with the giving up of the server
is the immediate choosing of the next user of the server,
without the releasing entity having yet become a
contender for the server. (2) The choosing of the next
user of the server is deferred until the releasing entity has
become a contender. (3) Neither of the above; that is,

without paying heed to other contenders, the releasing

entity recaptures the server immediately.
SIMAN implements (1) by default.
ProModel
implements (2). GPSS/H and AutoMod implement (3)
by default. In SLX, using a low-level Control Variable
as the resource state, the result is also (3). (However
developers could implement higher-level resource
constructs in SLX that behave in any of the three ways.)
8.3 The First in Line is Still Delayed
Suppose two Condition-Delayed entities are waiting in a
list because no units of a particular resource are idle.
Suppose the first entity needs two units of the resource,
whereas the second entity only needs one unit. Now
assume that one unit of the resource becomes idle. The
needs of the first list entity cannot yet be satisfied, but
the needs of the second entity can. What will happen?
There are at least three possible alternatives: (1)
Neither entity claims the idle resource unit. (2) The first
entity claims the one idle resource unit and waits for a
second unit. (3) The second entity claims the idle
resource unit and goes on its way.
As in Section 8.2, each of these alternatives comes into
play in the tools considered here. SIMAN (SEIZE) and
ProModel (GET or USE) implement (1) and (2)
respectively by default. AutoMod (GET or USE),
GPSS/H (ENTER or TEST), and SLX (WAIT UNTIL
on a Control Variable) implement (3) by default.
8.4 Yielding Control
Suppose the active entity wants to give control to one or
more Ready-State entities, but then needs to become the
active entity again before the simulation clock has been
advanced. This might be useful, for example, if the
active entity has opened a switch permitting a set of other
entities to move past a point in the model, and then needs
to re-close the switch after the forward movement has
been accomplished. (Perhaps a group of identicallyflavored cartons of ice cream is to be transferred from an
accumulation point to a conveyor leading to a one-flavorper-box packing operation.)
In SIMAN and AutoMod, the effect can be
accomplished approximately with a DELAY (SIMAN)
or WAIT FOR (AutoMod) that puts the active entity
into a Time-Delayed State for an arbitrarily short but
non-zero simulated time.
In ProModel, WAIT 0 can be used to put the active
entity back on the FEL. It will be returned later (at the
same simulated time) by the CUP to the Active State.
In GPSS/H, the active Transaction (Xact) can
execute a YIELD (BUFFER) Block to shift from the

Inside Discrete-Event Simulation Software: How it Works and Why it Matters

Active State to the Ready State and restart the CEC scan.
Higher-priority (and higher-ranked same priority) Xacts
on the CEC can then try to become active, one by one,
before the control-yielding Xact itself again becomes
active at the same simulated time. (A PRIORITY
PR,YIELD Block can alternatively be used in order to
reposition the just-active Xact behind equal-priority
Xacts on the CEC prior to restarting the scan.)
In SLX there is also a YIELD statement. Anormal
YIELD shifts the active Puck to the back of its priority
class on the CEC and picks up the next Puck. It is also
possible to YIELD to a specific other Puck that is on the
CEC, in which case the active Puck is not shifted.
8.5 Conditions Involving the Clock
If an entity needs to wait until a particular clock value
has been reached, every language has a time-dela y for
FEL waiting. But what if an entity needs to wait for a
compound condition involving the clock, such as wait
until my input buffer is empty or it is exactly 5:00 PM?
A typical approach to this is to clone a dummy
(shadow) entity to do the time-based waiting.
Management of dummy entities can be cumbersome,
particularly for very complex rules. ProModel has no
polled waiting, so a dummy entity would be required.
If a single entity tries to wait on a compound condition
involving the clock, other problems can arise. SIMAN
and AutoMod detect the truth of these conditions through
their end-of-EMP polling mechanisms. GPSS/H also
detects the truth through its version of polled waiting
(refusal-mode TEST). But in the absence of a clone that
waits on the FEL until exactly 5:00 PM, all three of those
tools are subject to the possibility that the first EMP that
finds the condition true has a clock value greater than
5:00 PM.
SLX recognizes the clock as a related wait-until target.
A WAIT UNTIL using a future clock value in a way that
contributes to the falseness of the condition will cause
the Puck to be scheduled onto the FEL to force an EMP
at the precise time referenced. This solves the greaterthan-the-desired-time problem. Note that this Puck may
also be waiting on one or more delay lists.
8.6 Mixed-Mode Waiting
Suppose many entities are waiting to capture a
particular resource, while a user-defined controller entity
is waiting for the condition shift status is off-shift and
number waiting is less than six and resource is not
currently in use to take some action (such as shutting the
resource down, in languages that allow user-defined
entities to shut down resources, or printing a status
message). How can we guarantee that the controller will

21

be able to cut in front of the waiting entities at the

appropriate instant (before the resource is recaptured)?
One way to handle this would be through entity
priorities, in languages that offer this mechanism.
However, as described below, that may not work even if
the controller has higher priority than any other entity.
The key issue is the method used to implement the
waiting. If it is related for the entities and polled for the
compound condition, things can get complicated. (This
is what we mean by the term mixed-mode waiting.)
Every time the resource comes free, a new entity will be
selected from a delay list immediately in SIMAN and via
the CEL in AutoMod, in both cases preceding the end-ofEMP checking for polled wait conditions (and thereby
ignoring the entity priority of the controller). There are
many ways to work around this if desired, such as using a
different type of operation to force a polled wait for
entities wishing to use the resource.
In GPSS/H, using a high-priority controller
Transaction at a refusal-mode TEST Block, the
controller waits at the front of the CEC. The RELEASE
of the Facility will trigger a scan restart and the
controller will do its job.
In ProModel there is no polled waiting but there can
be related waiting on compound conditions involving
Variables. Variables would have to be defined and
manipulated for each element of the Boolean condition
and, to assure equal competition, the entities might also
have to use WAIT UNTIL instead of GET or USE.
Another possibility with ProModel would be to have the
entity that frees the resource do some state-checking
right away (in effect becoming a surrogate for the
controller) . This is possible because of the deferredselection method used by ProModel (see Section 8.2).
In the related waiting of SLX, a Puck awaiting a
compound condition will be registered on the delay lists
of those (and only those) Control Variables that are
contributing to the falseness of the condition at the time
it is evaluated. The SLX architecture (in which only
global or local Control Variables and the clock can be
referenced in any sort of conditional wait at the lowest
level) assures that there will already be Variables
underlying the state changes being monitored. The
model developer needs only to be sure they are defined
as Control Variables.
8.7 Interactive Model Verification
We now comment briefly on why a detailed understanding of how simulation software works supports interactive probing of simulation-model behavior.
In general, simulation models can be run interactively
or in batch mode. Interactive runs are of use in checking
out (verifying) model logic during model-building and in

22

Schriber and Brunner

troubleshooting a model when execution errors occur.

Batch mode is then used to make production runs.
Interactive runs put a magnifying glass on a simulation
model while it executes. The modeler can follow the active entity step by step and display the current and future
events lists and the delay and user-managed lists as well
as other aspects of the model. These activities yield
valuable insights into model behavior for the modeler
who knows the underlying concepts. Without such
knowledge, the modeler might not take full advantage of
the interactive tools provided by the software or, worse
yet, might even avoid using the tools.
ACKNOWLEDGMENTS
Much of the information in this paper was derived from
conversations with software-vendor personnel. The
authors gratefully acknowledge the support provided by
David T. Sturrock, Deborah A. Sadowski, C. Dennis
Pegden and Vivek Bapat, all of Systems Modeling
Corporation; Charles Harrell of ProModel Corporation;
Kenneth Farnsworth and Tyler Phillips, of
AutoSimulations, Inc.; and Robert C. Crain and James O.
Henriksen, of Wolverine Software Corporation.
REFERENCES
Balci, O. 1988. The implementation of four conceptual
frameworks for simulation modeling in high-level
languages. In Proceedings of the 1988 Winter
Simulation Conference, ed. M. Abrams, P. Haigh, and
J. Comfort, 287-295. La Jolla, California: Society for
Computer Simulation.
Crain, R. C. 1996. Simulation using GPSS/H. In
Proceedings of the 1996 Winter Simulation
Conference, ed. J. Charnes, D. Morrice, D. Brunner,
and J. Swain, 453-459. Piscataway, New Jersey:
Institute of Electrical and Electronics Engineers.
Henriksen, J. O. 1996. An introduction to SLX. In
Proceedings of the 1996 Winter Simulation
Conference, ed. J. Charnes, D. Morrice, D. Brunner,
and J. Swain, 468-475. Piscataway, New Jersey:
Institute of Electrical and Electronics Engineers.
Henriksen, J. O., and R. C. Crain. 1998. GPSS/H
Reference Manual, 4th Edition. Annandale, Virginia:
Wolverine Software Corporation.
Pegden, C. D., R. E. Shannon, and R. P. Sadowski.
1995. Introduction to Simulation Using SIMAN,
Second Edition. New York, New York: McGraw-Hill.
Phillips, T. 1997. Know your AutoMod current events.
In AutoFlash newsletter, July 1997, Volume 10,
Number 7. Bountiful, Utah: AutoSimulations, Inc.
ProModel Corporation. 1995. ProModel Version 3
Users Guide. Orem, Utah: ProModel Corporation.

Schriber, T. J. 1991. An Introduction to Simulation

Using GPSS/H. New York, New York: John Wiley &
Sons.
Schriber, T. J., and D. T. Brunner. 1996. Inside
simulation software: How it works and why it matters.
In Proceedings of the 1996 Winter Simulation
Conference, ed. J. Charnes, D. Morrice, D. Brunner,
and J. Swain, 23-30. Piscataway, New Jersey: Institute
of Electrical and Electronics Engineers.
Schriber, T. J., and D. T. Brunner. 1998. How discreteevent simulation software works. In The Handbook of
Simulation, ed. J. Banks. New York, New York: John
Wiley & Sons.
Swain, J. J. 1995. Simulation survey: Tools for process
understanding and improvement.
OR/MS Today,
August 95, 64-79, Baltimore, Maryland: INFORMS.
Wolverine Software Corporation. 1996. SLX: An
introduction for GPSS/H users. Annandale, Virginia:
Wolverine Software Corporation.
AUTHOR BIOGRAPHIES
DANIEL T. BRUNNER is President of Systemflow
Simulations, Inc., a services firm active in manufacturing, material handling, distribution, transportation, health
care, computer systems, and mining. He received a
B.S.E.E. from Purdue University and an MBA from The
University of Michigan. He has served as Winter
Simulation Conference Publicity Chair (1988), Business
Chair (1992), and General Chair (1996). He is a member
of IIE and SCS.
THOMAS J. SCHRIBER is a Professor of Computer
and Information Systems at The University of Michigan.
He is a Fellow of the Institute of Decision Sciences and
is the 1996 recipient of the INFORMS College of
Simulation Distinguished Service Award. He teaches
modeling and decision analysis and discrete-event
simulation in Michigans MBA program while doing research and consulting in simulation. He is a member of
ASIM (the German-language simulation society), DSI,
IIE, and INFORMS.