Executing Chapel Programs ¶

Once you have compiled a Chapel program using the chpl compiler, you can execute it from the command-line like any other program. This document discusses command line options, environment variables, and other topics having to do with running programs written in Chapel.

Contents

Executing Chapel Programs

Getting Help ¶

Using the -h or --help flags will print out help for the executable. For example:

./myprogram --help

This flag lists all of the standard flags that can be used with a Chapel program, as well as a list of the configuration variables defined by the program and their types. If the configuration variable has been set on the command line, its value is also shown.

Setting Configuration Variables ¶

Command Line ¶

Configuration constants and variables defined in a Chapel program can have their default values overridden on the command line using the -s or -- flags. Either flag takes the name of the configuration variable followed by an equals character (=) and the value to assign to it. This value must be a legal Chapel literal for the type of the variable. (Exception: for a string literal, the surrounding quotes are implicit.) In our current implementation, no extra spaces may appear between these elements. Thus, the general form is:

./myprogram --<cfgVar>=<val>

or:

./myprogram -s<cfgVar>=<val>

As an example, compile the hello2-module.chpl example which prints a user-definable message:

chpl -o hello2-module $CHPL_HOME/examples/hello2-module.chpl

This program defines a configuration constant, message, indicating the string to print to the console, set to "Hello, world!" by default. To override the default, you can use:

./hello2-module --message='Hello, Chapel users!'

or:

./hello2-module -smessage='Hello, Chapel users!'

The default value of a configuration constant or variable can also be overridden at compilation time with a -s option. (The surrounding quotes must be provided for a string literal.) For example:

cd $CHPL_HOME/examples
chpl -smessage='"Hello from the Chapel compiler"' hello2-module.chpl

or:

chpl -s message="'Hello from the Chapel compiler'" hello2-module.chpl

The compiler-established default can still be overridden when executing the program, as shown above.

Command-line overrides of configs may also use fully qualified names. Thus, each of the above examples could have referred to message as Hello.message instead. This is useful for disambiguating when multiple modules declare configs with the same name. In addition, private config declarations must be set using fully-qualified names.

Configuration File ¶

Configuration values can also be passed to a Chapel program through a configuration file, specified by the execution time -f option. Configuration files can contain a whitespace- or newline-delimited list of keys and values separated by an assignment operator =. Comments begin with the # character. The examples below demonstrate this format.

Consider the following program:

// program.chpl
config const msg: string,
             val1: real,
             val2: real;


proc main() {
  writeln(msg);
  writeln(val1);
  writeln(val2);
}

The above program can have its configuration variables defined by this configuration file:

# program.input

msg="hello world"
val1=1.61803
val2=3.14159

Configuration files can contain a whitespace- or newline-separated list of configuration assignments and comments are supported with the # character. The configuration file above can also be written like this:

# program.input

val1=1.61803 val2=3.14159
msg="hello world" # This is a comment

The program.input is passed during execution with the -f flag:

# config variables are populated by program.input values
./program -fprogram.input

Warning

Assignments cannot contain whitespaces outside of quotes, so the following configuration file would result in an error:

# bad.input

# The additional whitespace will result in an error
val1 = 1.161803

Non-Configuration Variable Arguments ¶

Chapel programs can also accept C-like command line arguments to their main() procedure in addition to the aforementioned configuration variables. See main() Functions for more information.

Setting the Number of Locales ¶

For multi-locale Chapel executions, the number of locales on which to execute a program is specified on the executable’s command-line. This can be set either using the -nl flag, or by assigning to the built-in numLocales configuration constant using the normal mechanisms. So, to execute on four locales, one could use:

./myprogram -nl 4

or:

./myprogram --numLocales=4

or:

./myprogram -snumLocales=4

For users running with CHPL_COMM=none (the default), only one locale can be used. See Multilocale Chapel Execution for more information about executing on multiple locales.

Multi-locale programs often use a launcher executable to do some initial command-line checking before spawning the real program, which is then stored in a second binary named <original_binary_name>_real. See Chapel Launchers for more information about the launcher executable.

Controlling Degree of Data Parallelism ¶

Data parallel operations over ranges, default domains, and default arrays permit the number of tasks used to implement the data parallelism to be specified using the following built-in configuration constants:

dataParTasksPerLocale
Number of Chapel tasks to use to execute forall loops (default: number of physical CPUs on the node, which may be reduced by CHPL_RT_NUM_THREADS_PER_LOCALE. When CHPL_TASKS=qthreads, CHPL_RT_NUM_THREADS_PER_LOCALE can also increase this, up to the number of logical CPUs).

dataParIgnoreRunningTasks
If true, always use dataParTasksPerLocale tasks to execute forall loops. If false, reduce the number of tasks used by the number of tasks already running (default: false).

dataParMinGranularity
The number of tasks used to execute forall loops should be reduced such that the number of iterations per task is never less than the specified value (default: 1).

Most Chapel standard distributions also use identically named constructor arguments to control the degree of data parallelism within each locale when iterating over its domains and arrays. The default values for these arguments are set to the corresponding global configuration constants.

Environment Variable Control over Chapel Behavior ¶

Chapel uses environment variables to control the number of threads used at execution time and the call stack size, among other things. In many cases third-party packages used by Chapel define their own environment variables to provide the same or similar control. When this is the case, the Chapel environment variable has precedence over the third-party package environment variable, which in turn has precedence over the Chapel default.

As an example, with CHPL_TASKS=qthreads, Chapel and Qthreads both have environment variables that can be used to set the task stack size. In this case the Chapel CHPL_RT_CALL_STACK_SIZE environment variable will override the Qthreads QT_STACK_SIZE environment variable if both are set. However, if only QT_STACK_SIZE is set it will override the Chapel default call stack size.

The Chapel environment variables that control execution time behavior are as follows:

CHPL_RT_CALL_STACK_SIZE
size of the call stack for a task

CHPL_RT_MAX_HEAP_SIZE
per-locale size of the heap used for dynamic allocation in multilocale programs

CHPL_RT_NUM_THREADS_PER_LOCALE
number of threads used to execute tasks

CHPL_RT_USE_PU_KIND
which kind of processing units to use on CPUs with hybrid processing units. Values are “performance”, “efficiency”, or “all”.

There is a bit more information on CHPL_RT_CALL_STACK_SIZE, CHPL_RT_NUM_THREADS_PER_LOCALE, and CHPL_RT_USE_PU_KIND below, and more detailed discussions in Chapel Tasks and Using Chapel on Cray Systems.

Controlling the Call Stack Size ¶

The main Chapel program requires space for its call stack, as does any task created by it. This stack space has a fixed size. It is created automatically when the program or task starts executing, and remains in existence until it completes. The default call stack size is ~8 MiB on Linux-based systems, since this is a common value for the process stack limit on such systems. On Cygwin systems the default call stack size is ~2 MiB. Note that up to 4 system pages of each stack may be reserved for use by the tasking layer. Up to 2 pages for runtime data structures and up to 2 additional pages if guard pages (–stack-checks) are enabled.

The default call stack size may not be appropriate in all cases. For programs in which some tasks have large stack frames or deep call trees it may be too small, leading to stack overflow. For programs which use tasks only for data parallelism it may be unnecessarily large. Stacks that are unnecessarily large are typically only a problem for programs in which many tasks (thus their stacks) exist at once, when using a comm layer that has to pre-register memory. For the particular case of using the native runtime communication layers on Cray XC and HPE Cray EX systems, further discussion about this can be found in Using Chapel on Cray Systems.

The following environment variable can be used to change the task call stack size.

CHPL_RT_CALL_STACK_SIZE
Size of the call stack for a task. A plain numeric value indicates bytes. A suffix can be appended to indicate larger units:

k, K: KiB (2**10 bytes)

m, M: MiB (2**20 bytes)

g, G: GiB (2**30 bytes)

Controlling the Number of Threads ¶

The following environment variable can be used to change the number of system threads used by a program.

CHPL_RT_NUM_THREADS_PER_LOCALE
Controls the number of threads used on each locale when running the program.

See Chapel Tasks for more information on the role of this variable in creating threads and executing tasks for the various tasking layers.

Controlling the Kind of Processing Units ¶

Some CPUs, such as Intel’s “Alder Lake” family of processors, have hybrid processing units, some of which are “performance” and others of which are “efficiency”. The following environment variable can be used to select the kind of processing units used by a program.

CHPL_RT_USE_PU_KIND
Specifies which kind of processing units to use on CPUs with hybrid processing units. Values are “performance”, “efficiency”, or “all”.

By default the Chapel runtime will only use “performance” processing units. Note that if set to “all” the runtime will run tasks on all available cores/processing units indiscriminately ; it will not take the difference in performance into account when assigning tasks to processing units.

This environment variable has no effect if the processor does not have hybrid processing units.

Controlling the Amount of Non-User Output ¶

The compiler-generated executable supports verbose and quiet modes that control the amount of Chapel-generated information printed by the executable.

-v, --verbose

Print more information. For example, print the launcher commands used to start the program (if any) and print a message from each locale when the program starts executing there.

-q, --quiet

Print less information. For example, suppress run-time warnings that are printed by default.

Oversubscription ¶

In multi-locale Chapel executions programs can run “oversubscribed”, with more than one Chapel locale per system compute node. Both the gasnet and ofi communication layers support this. However, oversubscription can cause serious performance degradation due to resource contention, when multiple Chapel locales (program instances) all proceed as if the resources of the entire compute node are theirs to use. As a result, on a single system node, a program will almost always run faster with just a single locale than it will with multiple locales. Nevertheless, sometimes oversubscription is useful, such as for testing multilocale Chapel functionality when multiple system nodes are not actually available.

As a partial workaround for the resource contention problem, setting the following environment variable often improves performance when running oversubscribed:

export CHPL_RT_OVERSUBSCRIBED=yes

This causes various software components, from launchers to the runtime, to be more considerate in how they use node resources.

Stack Traces ¶

When compiled with unwind support (CHPL_UNWIND=bundled or CHPL_UNWIND=system), Chapel can print a stacktrace when it halts due to an error. For example, when compiling and running the following program:

// outofbounds.chpl
var A: [1..3] real;
A[5] = 1.0;

./outofbounds

outofbounds.chpl:2: error: halt reached - array index out of bounds
note: index was 5 but array bounds are 1..3
Stacktrace

halt() at $CHPL_HOME/modules/standard/Errors.chpl:741
checkAccess() at $CHPL_HOME/modules/internal/ChapelArray.chpl:2675
chpl__init_outofbounds() at outofbounds.chpl:1

The stacktrace will be printed. This behavior can also be disabled at runtime with the environment variable

export CHPL_RT_UNWIND=no

Note that compiling programs with --fast effectively forces CHPL_RT_UNWIND=no so that no stacktrace will be printed even if the Chapel compiler was built with unwind support. And in the above example, --fast also omits the bounds check.

Launcher Support ¶

For multilocale execution (see Multilocale Chapel Execution), Chapel programs are executed indirectly by a launcher. This section covers command line options that assist launchers in doing their job. These options are not supported for general use. We document them here so that their presence in, say, the output produced by -v or --dry-run can be understood.

At present there is only one launcher support option:

-E <envVar=val>

set the given environment variable envVar to val.