# on-clauses: controlling locality/affinity¶

As mentioned previously, all Chapel programs begin execution as a single task running on locale #0. Tasks can execute on remote locales using on-clauses. An on-clause prefixes another statement, specifying where it should be executed. This can be thought of as migrating the task that encountered the on-clause to a (potentially) remote locale. Once the on-clause’s statement completes, the original task continues executing on the original locale.

## Simple On-Clause Examples¶

As an example, the following program iterates sequentially through all of the Locales on which the program is running, conceptually migrating the main task to that locale, where it writes out a message indicating the locale’s ID:

examples/users-guide/locality/onClause.chpl
for loc in Locales do
on loc do
writeln("Hello from locale #", here.id);


Running this program on four locales generates:

Hello from locale #0
Hello from locale #1
Hello from locale #2
Hello from locale #3


Here’s a minor change to this example which demonstrates that the original task starts and ends on locale #0:

examples/users-guide/locality/onClausePlusOrig.chpl
for loc in Locales {
writeln("Started on locale #", here.id);
on loc do
writeln("Hello from locale #", here.id);
writeln("Back on    locale #", here.id, "\n");
}


When run on three locales, this program generates:

Started on locale #0
Hello from locale #0
Back on    locale #0

Started on locale #0
Hello from locale #1
Back on    locale #0

Started on locale #0
Hello from locale #2
Back on    locale #0


Note that the statement following an on-clause may include compound statements or function calls, so can result in an arbitrary amount of code being executed remotely.

## Locality is Orthogonal to Parallelism¶

It’s important to note that Chapel’s features for controlling locality are orthogonal to those used to create parallelism. Specifically, while both of the above programs run using multiple locales, neither is a parallel program. Instead, a single conceptual task is moving between locales as it executes. Similarly, all parallel programs from previous sections will only execute on a single locale—that is, they only result in shared-memory parallelism.

In order to get distributed memory parallelism, all that’s required is to compose Chapel’s features for parallelism and locality. As a simple example, the following program uses a coforall loop to create a task per locale in combination with an on-clause to execute each task on its corresponding locale:

examples/users-guide/locality/coforallPlusOn.chpl
coforall loc in Locales do
on loc do
writeln("Hello from locale ", loc.id, " of ", numLocales);


In effect, this creates an SPMD (Single Program, Multiple Data) style of parallelism. When running on four locales, the output will be a nondeterministic permutation of the following four lines:

Hello from locale 0 of 4
Hello from locale 1 of 4
Hello from locale 2 of 4
Hello from locale 3 of 4


## Data-driven On-clauses¶

While the on-clauses in the examples above use a locale value to specify where execution should take place, more generally, on-clauses can be specified using arbitrary expressions (typically variable references). In such cases, execution moves to the locale on which the variable or expression is stored. These are referred to as data-driven on-clauses.

As a simple example of a data-driven on-clause, the following program moves its task to the locale on which x is stored:

var x = 42;

Executing where x is stored, which is locale #0