Task Parallelism and Synchronization

Chapel supports both task parallelism and data parallelism. This chapter details task parallelism as follows:

Tasks and Task Parallelism

A Chapel task is a distinct context of execution that may be running concurrently with other tasks. Chapel provides a simple construct, the begin statement, to create tasks, introducing concurrency into a program in an unstructured way. In addition, Chapel introduces two type qualifiers, sync and single, for synchronization between tasks.

Chapel provides two constructs, the cobegin and coforall statements, to introduce concurrency in a more structured way. These constructs create multiple tasks but do not continue until these tasks have completed. In addition, Chapel provides two constructs, the sync and serial statements, to insert synchronization and suppress parallelism. All four of these constructs can be implemented through judicious uses of the unstructured task-parallel constructs described in the previous paragraph.

Tasks are considered to be created when execution reaches the start of a begin, cobegin, or coforall statement. When the tasks are actually executed depends on the Chapel implementation and run-time execution state.

A task is represented as a call to a task function, whose body contains the Chapel code for the task. Variables defined in outer scopes are considered to be passed into a task function by default intent, unless a different task intent is specified explicitly by a task-intent-clause.

Accesses to the same variable from different tasks are subject to the Memory Consistency Model (Memory Consistency Model). Such accesses can result from aliasing due to ref argument intents or task intents, among others.

The Begin Statement

The begin statement creates a task to execute a statement. The syntax for the begin statement is given by

begin-statement:
  'begin' task-intent-clause[OPT] statement

Control continues concurrently with the statement following the begin statement.

Example (beginUnordered.chpl).

The code

begin writeln("output from spawned task");
writeln("output from main task");

executes two writeln statements that output the strings to the terminal, but the ordering is purposely unspecified. There is no guarantee as to which statement will execute first. When the begin statement is executed, a new task is created that will execute the writeln statement within it. However, execution will continue immediately after task creation with the next statement.

A begin statement creates a single task function, whose body is the body of the begin statement. The handling of the outer variables within the task function and the role of task-intent-clause are defined in Task Intents.

Yield and return statements are not allowed in begin blocks. Break and continue statements may not be used to exit a begin block.

Synchronization Variables

Synchronization variables have a logical state associated with the value. The state of the variable is either full or empty. Normal reads of a synchronization variable cannot proceed until the variable’s state is full. Normal writes of a synchronization variable cannot proceed until the variable’s state is empty.

Chapel supports two types of synchronization variables: sync and single. Both types behave similarly, except that a single variable may only be written once. Consequently, when a sync variable is read, its state transitions to empty, whereas when a single variable is read, its state does not change. When either type of synchronization variable is written, its state transitions to full.

sync and single are type qualifiers and precede the type of the variable’s value in the declaration. sync and single are supported for the primitive types nothing, bool, int, uint, real, imag, complex, range, bytes, and stringPrimitive Types); for enumerated types ( Enumerated Types); and for class types (Class Types) and record types (Record Types). For sync variables of class type, the full/empty state applies to the reference to the class object, not to its member fields.

If a task attempts to read or write a synchronization variable that is not in the correct state, the task is suspended. When the variable transitions to the correct state, the task is resumed. If there are multiple tasks blocked waiting for the state transition:

  • for a sync variable, one task is non-deterministically selected to proceed and the others continue to wait

  • for a single variable, all tasks are selected to proceed.

A synchronization variable is specified with a sync or single type given by the following syntax:

sync-type:
  'sync' type-expression

single-type:
  'single' type-expression

A default-initialized synchronization variable will be empty. A synchronization variable initialized from another expression will be full and store the value from that expression.

Example (beginWithSyncVar.chpl).

The code

class Tree {
  var isLeaf: bool;
  var left, right: unmanaged Tree?;
  var value: int;

  proc sum():int {
    if (isLeaf) then
       return value;

    var x$: sync int;
    begin x$.writeEF(left!.sum());
    var y = right!.sum();
    return x$.readFE() + y;
  }
}

the sync variable x$ is assigned by an asynchronous task created with the begin statement. The task returning the sum waits on the reading of x$ until it has been assigned. By convention, synchronization variables end in $ to provide a visual cue to the programmer indicating that the task may block.

Example (singleVar.chpl).

The following code implements a simple split-phase barrier using a single variable.

var count$: sync int = n;  // counter which also serves as a lock
var release$: single bool; // barrier release

forall t in 1..n do begin {
  work(t);
  var myc = count$.readFE();  // read the count, set state to empty
  if myc!=1 {
    write(".");
    count$.writeEF(myc-1);   // update the count, set state to full
    // we could also do some work here before blocking
    release$.readFF();
  } else {
    release$.writeEF(true);  // last one here, release everyone
    writeln("done");
  }
}

In each iteration of the forall loop after the work is completed, the task reads the count$ variable, which is used to tally the number of tasks that have arrived. All tasks except the last task to arrive will block while trying to read the variable release$. The last task to arrive will write to release$, setting its state to full at which time all the other tasks can be unblocked and run.

If a formal argument with a default intent either has a synchronization type or the formal is generic (Formal Arguments of Generic Type) and the actual has a synchronization type, the actual must be an lvalue and is passed by reference. In these cases the formal itself is an lvalue, too. The actual argument is not read or written during argument passing; its state is not changed or waited on. The qualifier sync or single without the value type can be used to specify a generic formal argument that requires a sync or single actual.

When the actual argument is a sync or single and the corresponding formal has the actual’s base type or is implicitly converted from that type, a normal read of the actual is performed when the call is made, and the read value is passed to the formal.

Predefined Single and Sync Methods

The following methods are defined for variables of sync and single type.

proc (sync t).readFE(): t

Returns the value of the sync variable. This method blocks until the sync variable is full. The state of the sync variable is set to empty when this method completes. This method implements the normal read of a sync variable.

proc (sync t).readFF(): t
proc (single t).readFF(): t

Returns a copy of the value of the sync or single variable. This method blocks until the sync or single variable is full. The state of the sync or single variable remains full when this method completes. This method implements the normal read of a single variable.

proc (sync t).readXX(): t
proc (single t).readXX(): t

This method does not block and the state of the sync or single variable is unchanged when this method completes.

This function returns:

  • for a full sync or single, a copy of the value stored

  • for an empty sync or single, the implementation will return either a new default-initialzed value of type t or the last value stored.

proc (sync t).writeEF(v: t)
proc (single t).writeEF(v: t)

Assigns v to the value of the sync or single variable. This method blocks until the sync or single variable is empty. The state of the sync or single variable is set to full when this method completes. This method implements the normal write of a sync or single variable.

proc (sync t).writeFF(v: t)

Assigns v to the value of the sync variable. This method blocks until the sync variable is full. The state of the sync variable remains full when this method completes.

proc (sync t).writeXF(v: t)

Assigns v to the value of the sync variable. This method is non-blocking and the state of the sync variable is set to full when this method completes.

proc (sync t).reset()

Assigns the default value of type t to the value of the sync variable. This method is non-blocking and the state of the sync variable is set to empty when this method completes.

proc (sync t).isFull: bool
proc (single t).isFull: bool

Returns true if the sync or single variable is full and false otherwise. This method is non-blocking and the state of the sync or single variable is unchanged when this method completes.

Atomic Variables

atomic is a type qualifier that precedes the variable’s type in the declaration. An atomic variable is specified with an atomic type given by the following syntax:

atomic-type:
  'atomic' type-expression

For example, the following code declares an atomic variable x that stores an int:

var x: atomic int;

Such an atomic variable that is declared without an initialization expression will store the default value of the contained type (i.e. 0 or false).

Atomic variables can also be declared with an initial value:

var y: atomic int = 1;

Chapel currently supports atomic operations for bools, all supported sizes of signed and unsigned integers, as well as all supported sizes of reals. Note that not all operations are supported for all atomic types. The supported types are listed for each operation.

Rationale.

The choice of supported atomic variable types as well as the atomic operations was strongly influenced by the C11 standard.

Most atomic methods accept an optional argument named order of type memoryOrder. The order argument is used to specify the ordering constraints of atomic operations. The supported memoryOrder values are:

  • memoryOrder.relaxed

  • memoryOrder.acquire

  • memoryOrder.release

  • memoryOrder.acqRel

  • memoryOrder.seqCst

See also Memory Consistency Model and in particular Non-Sequentially Consistent Atomic Operations for more information on the meaning of these memory orders.

Unless specified, the default for the memoryOrder parameter is memoryOrder.seqCst.

Implementors’ note.

Not all architectures or implementations may support all memoryOrder values. In these cases, the implementation should default to a more conservative ordering than specified.

proc atomicFence(param order: memoryOrder = memoryOrder.seqCst)

An atomic fence that establishes an ordering of non-atomic and relaxed atomic operations.

atomic (bool)
proc read(param order: memoryOrder = memoryOrder.seqCst): bool
Returns

The stored value.

proc write(value: bool, param order: memoryOrder = memoryOrder.seqCst): void

Stores value as the new value.

proc exchange(value: bool, param order: memoryOrder = memoryOrder.seqCst): bool

Stores value as the new value and returns the original value.

proc compareExchange(ref expected: bool, desired: bool, param order: memoryOrder = memoryOrder.seqCst): bool

Stores desired as the new value, if and only if the original value is equal to expected. Returns true if desired was stored, otherwise updates expected to the original value.

proc compareExchange(ref expected: bool, desired: bool, param success: memoryOrder, param failure: memoryOrder): bool
proc compareExchangeWeak(ref expected: bool, desired: bool, param order: memoryOrder = memoryOrder.seqCst): bool

Similar to compareExchange, except that this function may return false even if the original value was equal to expected. This may happen if the value could not be updated atomically.

This weak version is allowed to spuriously fail, but when compareExchange is already in a loop, it can offer better performance on some platforms.

proc compareExchangeWeak(ref expected: bool, desired: bool, param success: memoryOrder, param failure: memoryOrder)
proc compareAndSwap(expected: bool, desired: bool, param order: memoryOrder = memoryOrder.seqCst): bool

Stores desired as the new value, if and only if the original value is equal to expected. Returns true if desired was stored.

proc testAndSet(param order: memoryOrder = memoryOrder.seqCst): bool

Stores true as the new value and returns the old value.

proc clear(param order: memoryOrder = memoryOrder.seqCst): void

Stores false as the new value.

proc waitFor(value: bool, param order: memoryOrder = memoryOrder.seqCst): void
Arguments

value – Value to compare against.

Waits until the stored value is equal to value. The implementation may yield the running task while waiting.

atomic (T)
proc read(param order: memoryOrder = memoryOrder.seqCst): T
Returns

The stored value.

proc write(value: T, param order: memoryOrder = memoryOrder.seqCst): void

Stores value as the new value.

proc exchange(value: T, param order: memoryOrder = memoryOrder.seqCst): T

Stores value as the new value and returns the original value.

proc compareExchange(ref expected: T, desired: T, param order: memoryOrder = memoryOrder.seqCst): bool

Stores desired as the new value, if and only if the original value is equal to expected. Returns true if desired was stored, otherwise updates expected to the original value.

proc compareExchange(ref expected: T, desired: T, param success: memoryOrder, param failure: memoryOrder): bool
proc compareExchangeWeak(ref expected: T, desired: T, param order: memoryOrder = memoryOrder.seqCst): bool

Similar to compareExchange, except that this function may return false even if the original value was equal to expected. This may happen if the value could not be updated atomically.

This weak version is allowed to spuriously fail, but when compareExchange is already in a loop, it can offer better performance on some platforms.

proc compareExchangeWeak(ref expected: T, desired: T, param success: memoryOrder, param failure: memoryOrder): bool
proc compareAndSwap(expected: T, desired: T, param order: memoryOrder = memoryOrder.seqCst): bool

Stores desired as the new value, if and only if the original value is equal to expected. Returns true if desired was stored.

proc fetchAdd(value: T, param order: memoryOrder = memoryOrder.seqCst): T
Returns

The original value.

Adds value to the original value and stores the result. Defined for integer and real atomic types.

proc add(value: T, param order: memoryOrder = memoryOrder.seqCst): void

Adds value to the original value and stores the result. Defined for integer and real atomic types.

proc fetchSub(value: T, param order: memoryOrder = memoryOrder.seqCst): T
Returns

The original value.

Subtracts value from the original value and stores the result. Defined for integer and real atomic types.

proc sub(value: T, param order: memoryOrder = memoryOrder.seqCst): void

Subtracts value from the original value and stores the result. Defined for integer and real atomic types.

proc fetchOr(value: T, param order: memoryOrder = memoryOrder.seqCst): T
Returns

The original value.

Applies the | operator to value and the original value, then stores the result.

Only defined for integer atomic types.

proc or(value: T, param order: memoryOrder = memoryOrder.seqCst): void

Applies the | operator to value and the original value, then stores the result.

Only defined for integer atomic types.

proc fetchAnd(value: T, param order: memoryOrder = memoryOrder.seqCst): T
Returns

The original value.

Applies the & operator to value and the original value, then stores the result.

Only defined for integer atomic types.

proc and(value: T, param order: memoryOrder = memoryOrder.seqCst): void

Applies the & operator to value and the original value, then stores the result.

Only defined for integer atomic types.

proc fetchXor(value: T, param order: memoryOrder = memoryOrder.seqCst): T
Returns

The original value.

Applies the ^ operator to value and the original value, then stores the result.

Only defined for integer atomic types.

proc xor(value: T, param order: memoryOrder = memoryOrder.seqCst): void

Applies the ^ operator to value and the original value, then stores the result.

Only defined for integer atomic types.

proc waitFor(value: T, param order: memoryOrder = memoryOrder.seqCst): void

Waits until the stored value is equal to value. The implementation may yield the running task while waiting.

proc type AtomicBool.=(ref a: AtomicBool, const ref b: AtomicBool)

Equivalent to a.write(b.read())

proc type AtomicT.=(ref a: AtomicT, const ref b: AtomicT)

Equivalent to a.write(b.read())

The Cobegin Statement

The cobegin statement is used to introduce concurrency within a block. The cobegin statement syntax is

cobegin-statement:
  'cobegin' task-intent-clause[OPT] block-statement

A new task and a corresponding task function are created for each statement in the block-statement. Control continues when all of the tasks have finished. The handling of the outer variables within each task function and the role of task-intent-clause are defined in Task Intents.

Return statements are not allowed in cobegin blocks. Yield statement may only be lexically enclosed in cobegin blocks in parallel iterators (Parallel Iterators). Break and continue statements may not be used to exit a cobegin block.

Example (cobeginAndEquivalent.chpl).

The cobegin statement

cobegin {
  stmt1();
  stmt2();
  stmt3();
}

is equivalent to the following code that uses only begin statements and single variables to introduce concurrency and synchronize:

var s1$, s2$, s3$: single bool;
begin { stmt1(); s1$.writeEF(true); }
begin { stmt2(); s2$.writeEF(true); }
begin { stmt3(); s3$.writeEF(true); }
s1$.readFF(); s2$.readFF(); s3$.readFF();

Each begin statement is executed concurrently but control does not continue past the final line above until each of the single variables is written, thereby ensuring that each of the functions has finished.

The Coforall Loop

The coforall loop is a variant of the cobegin statement in loop form. The syntax for the coforall loop is given by

coforall-statement:
  'coforall' index-var-declaration 'in' iteratable-expression task-intent-clause[OPT] 'do' statement
  'coforall' index-var-declaration 'in' iteratable-expression task-intent-clause[OPT] block-statement
  'coforall' iteratable-expression task-intent-clause[OPT] 'do' statement
  'coforall' iteratable-expression task-intent-clause[OPT] block-statement

The coforall loop creates a separate task for each iteration of the loop. Control continues with the statement following the coforall loop after all tasks corresponding to the iterations of the loop have completed.

The single task function created for a coforall and invoked by each task contains the loop body. The handling of the outer variables within the task function and the role of task-intent-clause are defined in Task Intents.

Return statements are not allowed in coforall blocks. Yield statement may only be lexically enclosed in coforall blocks in parallel iterators (Parallel Iterators). Break and continue statements may not be used to exit a coforall block.

Example (coforallAndEquivalent.chpl).

The coforall statement

coforall i in iterator() {
  body();
}

is equivalent to the following code that uses only begin statements and sync and single variables to introduce concurrency and synchronize:

var runningCount$: sync int = 1;
var finished$: single bool;
for i in iterator() {
  runningCount$.writeEF(runningCount$.readFE() + 1);
  begin {
    body();
    var tmp = runningCount$.readFE();
    runningCount$.writeEF(tmp-1);
    if tmp == 1 then finished$.writeEF(true);
  }
}
var tmp = runningCount$.readFE();
runningCount$.writeEF(tmp-1);
if tmp == 1 then finished$.writeEF(true);
finished$.readFF();

Each call to body() executes concurrently because it is in a begin statement. The sync variable runningCount$ is used to keep track of the number of executing tasks plus one for the main task. When this variable reaches zero, the single variable finished$ is used to signal that all of the tasks have completed. Thus control does not continue past the last line until all of the tasks have completed.

Task Intents

If a variable is referenced within the lexical scope of a begin, cobegin, or coforall statement and is declared outside that statement, it is subject to task intents. That is, this outer variable is considered to be passed as an actual argument to the corresponding task function at task creation time. All references to the variable within the task function implicitly refer to a shadow variable, i.e. the task function’s corresponding formal argument.

When the task construct is inside a method on a record and accesses a field of this, the field itself is treated as an outer variable. That is, it is passed as an actual argument to the task function and all references to the field within the task function implicitly refer to the corresponding shadow variable.

Each formal argument of a task function has the default argument intent by default. For variables of primitive and class types, this has the effect of capturing the value of the variable at task creation time and referencing that value instead of the original variable within the lexical scope of the task construct.

A formal can be given another argument intent explicitly by listing it with that intent in the optional task-intent-clause. For example, for variables of most types, the ref intent allows the task construct to modify the corresponding original variable or to read its updated value after concurrent modifications.

The syntax of the task intent clause is:

task-intent-clause:
  'with' ( task-intent-list )

task-intent-list:
  task-intent-item
  task-intent-item, task-intent-list

task-intent-item:
  formal-intent identifier
  reduce-scan-operator 'reduce' identifier
  class-type 'reduce' identifier
  task-private-var-decl

where the following intents can be used as a formal-intent: ref, in, const, const in, const ref. task-private-var-decl is defined in Task-Private Variables.

The reduce task intent specifies a reduction into the outer variable, which is provided to the right of the reduce keyword. The reduction operator is specified by either the reduce-scan-operator or the class-type in the same way as for a Reduction Expressions (see Reduction Expressions). At the start of each task the corresponding shadow variable is initialized to the identity value of the reduction operator. Within the task it behaves as a regular variable. In addition, it can be the left-hand side of the reduce= operator, which accumulates its right-hand side onto the shadow variable. At the end of each task its shadow variable is combined into the outer variable.

Open issue.

How should reduce task intent be defined for begin tasks? A reduction is legal only when the task completes before the program has exited the dynamic scope of the outer variable.

Reduce intents are currently work-in-progress. See also Reduce Intents technical note.

The implicit treatment of outer scope variables as the task function’s formal arguments applies to both module level and local variables. It applies to variable references within the lexical scope of a task construct, but does not extend to its dynamic scope, i.e., to the functions called from the task(s) but declared outside of the lexical scope. The loop index variables of a coforall statement are not subject to such treatment within that statement; however, they are subject to such treatment within nested task constructs, if any.

Rationale.

The primary motivation for task intents is to avoid some races on scalar/record variables, which are possible when one task modifies a variable and another task reads it. Without task intents, for example, it would be easy to introduce and overlook a bug illustrated by this simplified example:

{
  var i = 0;
  while i < 10 {
    begin {
      f(i);
    }
    i += 1;
  }
}

If all the tasks created by the begin statement start executing only after the while loop completes, and i within the begin is treated as a reference to the original i, there will be ten tasks executing f(10). However, the user most likely intended to generate ten tasks executing f(0), f(1), …, f(9). Task intents ensure that, regardless of the timing of task execution.

Another motivation for task intents is that referring to a captured copy in a task is often more efficient than referring to the original variable. That’s because the copy is a local constant, e.g. it could be placed in a register when it fits. Without task intents, references to the original variable would need to be implemented using a pointer dereference. This is less efficient and can hinder optimizations in the surrounding code, for example loop-invariant code motion.

Furthermore, in the above example the scope where i is declared may exit before all the ten tasks complete. Without task intents, the user would have to protect i to make sure its lexical scope doesn’t exit before the tasks referencing it complete.

We decided to treat cobegin and coforall statements the same way as begin. This is for consistency and to make the race-avoidance benefit available to more code.

We decided to apply task intents to module level variables, in addition to local variables. Again, this is for consistency. One could view module level variables differently than local variables (e.g. a module level variable is “always available”), but we favored consistency over such an approach.

We decided not to apply task intents to “closure” variables, i.e., the variables in the dynamic scope of a task construct. This is to keep this feature manageable, so that all variables subject to task intents can be obtained by examining just the lexical scope of the task construct. In general, the set of closure variables can be hard to determine, unwieldy to implement and reason about, it is unclear what to do with extern functions, etc.

We do not provide inout or out as task intents because they will necessarily create a data race in a cobegin or coforall. type and param intents are not available either as they do not seem useful as task intents.

Note

Future.

For a given intent, we would also like to provide a blanket clause, which would apply the intent to all variables. An example of syntax for a blanket ref intent would be ref *.

The Sync Statement

The sync statement acts as a join of all dynamically encountered begins from within a statement. The syntax for the sync statement is given by

sync-statement:
  'sync' statement
  'sync' block-statement

Return statements are not allowed in sync statement blocks. Yield statement may only be lexically enclosed in sync statement blocks in parallel iterators (Parallel Iterators). Break and continue statements may not be used to exit a sync statement block.

Example (syncStmt1.chpl).

The sync statement can be used to wait for many dynamically created tasks.

sync for i in 1..n do begin work();

The for loop is within a sync statement and thus the tasks created in each iteration of the loop must complete before the continuing past the sync statement.

Example (syncStmt2.chpl).

The sync statement

sync {
  begin stmt1();
  begin stmt2();
}

is similar to the following cobegin statement

cobegin {
  stmt1();
  stmt2();
}

except that if begin statements are dynamically encountered when stmt1() or stmt2() are executed, then the former code will wait for these begin statements to complete whereas the latter code will not.

The Serial Statement

The serial statement can be used to dynamically disable parallelism. The syntax is:

serial-statement:
  'serial' expression[OPT] 'do' statement
  'serial' expression[OPT] block-statement

where the optional expression evaluates to a boolean value. If the expression is omitted, it is as though ’true’ were specified. Whatever the expression’s value, the statement following it is evaluated. If the expression is true, any dynamically encountered code that would normally create new tasks within the statement is instead executed by the original task without creating any new ones. In effect, execution is serialized. If the expression is false, code within the statement will generates task according to normal Chapel rules.

Example (serialStmt1.chpl).

In the code

proc f(i) {
  serial i<13 {
    cobegin {
      work(i);
      work(i);
    }
  }
}

for i in lo..hi {
  f(i);
}

the serial statement in procedure f() inhibits concurrent execution of work() if the variable i is less than 13.

Example (serialStmt2.chpl).

The code

serial {
  begin stmt1();
  cobegin {
    stmt2();
    stmt3();
  }
  coforall i in 1..n do stmt4();
}

is equivalent to

stmt1();
{
  stmt2();
  stmt3();
}
for i in 1..n do stmt4();

because the expression evaluated to determine whether to serialize always evaluates to true.

Atomic Statements

Open issue.

This section describes a feature that is a work-in-progress. We seek feedback and collaboration in this area from the broader community.

The atomic statement is used to specify that a statement should appear to execute atomically from other tasks’ point of view. In particular, no task will see memory in a state that would reflect that the atomic statement had begun executing but had not yet completed.

Open issue.

This definition of the atomic statement provides a notion of strong atomicity since the action will appear atomic to any task at any point in its execution. For performance reasons, it could be more practical to support weak atomicity in which the statement’s atomicity is only guaranteed with respect to other atomic statements. We may also pursue using atomic type qualifiers as a means of marking data that should be accessed atomically inside or outside an atomic section.

The syntax for the atomic statement is given by:

atomic-statement:
  'atomic' statement

Example.

The following code illustrates the use of an atomic statement to perform an insertion into a doubly-linked list:

proc Node.insertAfter(newNode: Node) {
  atomic {
    newNode.prev = this;
    newNode.next = this.next;
    if this.next then this.next.prev = newNode;
    this.next = newNode;
  }
}

The use of the atomic statement in this routine prevents other tasks from viewing the list in a partially-updated state in which the pointers might not be self-consistent.