simcode

Discrete-event simulation programs in C, written by Rob Jones for CSci 726 (Discrete-Event Simulation) in 1997 and revised through 2002. The centerpiece is qnsim, a general-purpose simulator for networks of single-server queues. Around it sit the course exercises and the Park & Geyer random number and statistics library they all share.

In 2026 the code was brought up to C23 with a CMake build, clang-tidy, sanitizers and a regression suite. Every program that ran before produces byte-identical output.

Contents

Area	What it is
qnsim (qnet.c, qnsim*.c, queue.c)	Queuing network simulator: describe a network in a small C file, get steady-state or transient estimates with confidence intervals
rngs.c, rvgs.c, rvms.c	Park & Geyer library: 256-stream Lehmer random number generator, random variate generators, and pdf/cdf/inverse-cdf functions
ssq*.c, msq.c, mm1.c, sssim.c, tsim*.c, simjobs.c, mobile-server.c	Single- and multi-server queue simulations
sis*.c	(s, S) inventory system simulations
uvs.c, bvs.c, ddh.c, cdh.c, cdhcl.c, acs.c, estimate.c	Data analysis filters: statistics, histograms, autocorrelation and interval estimates for data read from stdin, one value per line
algC*.c, exC*.c, ex*.c, MMck.c, galileo.c, gamble*.c, buffon.c, monte*.c, urn*.c	Course exercises: generator tests, Monte Carlo experiments
qnsim/	Distributable copy of the simulator, kept identical to the root sources
tests/	Unit and regression test suite

Requirements

• A C23 compiler: Clang 18 or later, or GCC 14 or later. Apple clang 21 works.
• CMake 3.28 or later
• bash and perl for the regression tests
• Optional: Homebrew LLVM (brew install llvm) for clang-tidy and a current clang-format
• Optional: csh, only for the original run script

Quick start

make

This runs the dev CMake workflow: a debug build with AddressSanitizer and UndefinedBehaviorSanitizer, followed by the full test suite. Executables land in build/dev/bin/. Run the example network, the five-node network of example 9.5.4 in the course notes:

build/dev/bin/example954

Steady-State Simulation

  batch size.........:  1000
  number of batches..:  60
  confidence level...:  95%
  confidence interval:  +/- 5%

Expected number of jobs in the system:  14.3359 +/- 0.6757
Expected wait time of a job..........:  17.2721 +/- 0.8141

Other entry points:

Command	Does
make release	Optimized build and tests, in build/release/
make lint	Compiles everything through clang-tidy; any finding fails the build
make golden	Regenerates the regression golden files
make clean	Deletes all build trees

The make targets wrap CMake presets, so cmake --workflow --preset dev and friends work directly too. Every build uses -std=c23 -Wall -Wextra -Wpedantic -Wstrict-prototypes -Wmissing-prototypes -Wshadow -Werror.

qnsim, the queuing network simulator

Describing a network

A model is a C file that defines two functions, Network() and InitialState(). Node 0 is the outside world: jobs arrive from it and depart to it. Service nodes are numbered from 1.

#include "qnet.h"

static double arrive(void)  { return Exponential(2.0); }      /* mean inter-arrival time 2.0 */
static double service(void) { return Erlang(5, 2.0 / 7.0); }

void Network(void)
{
  Nodes(2);                      /* two service nodes, numbered 1 and 2      */
  Connect(0, 1, 1.0);            /* every arrival goes to node 1             */
  Connect(1, 2, 0.6);            /* node 1 routes to node 2 with prob 0.6    */
  Connect(1, 0, 0.4);            /*   ... or leaves the system with prob 0.4 */
  Connect(2, 0, 1.0);
  Arrival(arrive, 0.5);          /* function and mean arrival rate           */
  Service(1, service, 0.7);      /* function and mean service rate per node  */
  Service(2, service, 0.7);
}

void InitialState(void)
{
  /* empty network; or e.g. Initial(1, 3) to start node 1 with 3 jobs */
}

Call	Purpose
Nodes(n)	Declare n service nodes. Must come first.
Connect(i, j, p)	Route a job leaving node i to node j with probability p. Each row should sum to 1.
Arrival(f, rate)	External inter-arrival time function and its mean rate. Makes the network open.
Service(i, f, rate)	Service time function and mean service rate for node i.
Initial(i, j)	Start node i with j jobs. Call it from InitialState().

Any generator from rvgs.h can be used inside the time functions, for example Exponential, Erlang, Uniform, Normal or Triangle. Node i draws from random number stream i, so changing one node's distribution does not disturb the others.

Building a model

Add one line to CMakeLists.txt per variant:

simcode_qnsim_model(mymodel     mymodel.c STEADY    FIFO)
simcode_qnsim_model(mymodel-sjf mymodel.c TRANSIENT SJF  DEFINES PLOT)

Choice	Meaning
STEADY	qnsim-ss.c driver: batch means, adding batches of 1000 departures until the 95% interval is within 5% of the estimate. Reports mean jobs in the system and mean wait.
TRANSIENT	qnsim-ts.c driver: 10,000 replications from the initial state. Reports the mean number of external arrivals until flow in equals flow out.
FIFO / SJF	First-in first-out queues, or shortest job first with service times sampled on arrival
DEFINES PLOT	Transient driver prints per-replication data for plotting
DEFINES JACKSON	Model-defined switch. example954.c uses it to swap in exponential service.

The built-in targets are example954, example954-sjf, example954-jackson, example954-sjf-jackson, example954-ts, example954-ts-sjf, example954-ts-plot, model1 and model2.

The original csh wrapper still works for one-off builds: run -h lists its flags, and run example954 builds and runs a model with the same C23 flags.

Limitations

• Statistics are measured at node 0, so they are meaningful only for open networks. A closed network such as model2.c builds but never terminates.
• With FIFO queues, Initial(i, j) loads one job instead of j. See Known issues.

Tests

make                                   # or: ctest --preset dev
ctest --preset dev -L unit             # labels: unit, regress, sync
ctest --preset dev -R regress.ssq3     # a single case

• Unit tests (tests/unit/) cover the random number generator, every variate generator's mean and variance, the distribution functions against table values, the sort routine, both queue types, and the network engine.
• Regression tests run 52 programs and simulator variants with fixed seeds and compare their output byte-for-byte against tests/regress/golden/. Cases are listed in tests/regress/cases.txt.
• Sync check confirms that qnsim/ matches the root sources.

make golden regenerates the expected outputs after an intended change; review the diff before committing. Builds use -ffp-contract=off, so results are identical at every optimization level. The golden files were produced on macOS arm64; another libm may differ in the last digits.

What changed in 2026

Language. All sources compile as C23 with warnings as errors. bool is now the language keyword instead of a custom enum, all prototypes are explicit, helper functions are static, the FIFO queue macros became static inline functions, and C23 features such as nullptr, constexpr, [[noreturn]], [[maybe_unused]] and = {} are used where they make code clearer. Macros named INFINITY that clashed with <math.h> became NEVER.

Previously broken programs now build. MMck.c and monte_carlo.c had their recorded output pasted into the source; it is now a comment, and both programs reproduce that 1997 output exactly. sssim.c was C++ and is now C. mobile-server.c is reconnected to its circular-buffer queue. model1.c and model2.c were updated to the current Arrival/Service signatures.

Bugs fixed. None of these changed the output of any program that ran before.

• qnet.c: an out-of-bounds read in Init(), and an infinite loop in Initial() with SJF queues
• algC3.c: both of its arrays were allocated too small
• sssim.c, ex84x.c: the wait-time array overflowed by the last few arrivals; it now grows on demand
• simjobs.c: the running average started from an uninitialized value
• cdhcl.c: the argument parser read past the end of each option string
• circ_buf_queue.c: Take could not remove from a full queue
• 1-based arrays no longer step a pointer before the start of an allocation

Behavior changes.

• error() in the simulator now writes to stderr and exits with status 1. It used stdout and status 0.
• Programs that prompt for a seed exit with a message at end of input. They used to loop forever.
• The stdin filters stop at the first unreadable value instead of using garbage. cdhcl rejects unknown and missing options.

Known issues

• With FIFO queues, Initial(i, j) loads one job instead of j, because the FIFO Insert always adds one. The unit tests track this as a known bug.
• mobile-server.c has CHECKGRAPH defined, so it prints its routing graph and exits. Its first arrival is scheduled before the command-line arrival rate is read.
• ssq1.c and sis1.c need .dat input files that are not in the archive. buffon.c seeds from the clock, so it is not regression tested.
• multiproc.c is an unfinished sketch, and simcore.c and triangle.c are fragments without main. They are not built.

About the files

Compiled executables from the original SunOS (SPARC) and Cygwin builds are kept on disk but not tracked by git. Files such as qnet.c.1, qnsim.c.2 and makefile~, plus the .tar and .zip archives, are historical snapshots and are not built. The old makefiles (makefile.mci, makefile.qnsim, makefile.bak) and the go and .run scripts are superseded by CMake. The root makefile is now a wrapper around the CMake presets.

.clang-format describes the existing style for new code. The historical sources were deliberately not reformatted.

Credits

• Random number, variate and statistics library (rngs, rvgs, rvms, rng) and several textbook programs by Steve Park and Dave Geyer.
• Generator from S. K. Park and K. W. Miller, "Random Number Generators: Good Ones Are Hard To Find", Communications of the ACM, October 1988.
• qnsim, the exercises, and the revisions noted in the source headers by Rob Jones.