Package scala.collection.parallel

The base trait for all combiners. A combiner incremental collection construction just like a regular builder, but also implements an efficient merge operation of two builders via combine method. Once the collection is constructed, it may be obtained by invoking the result method.

The complexity of the combine method should be less than linear for best performance. The result method doesn't have to be a constant time operation, but may be performed in parallel.

A task support that uses an execution context to schedule tasks.

It can be used with the default execution context implementation in the scala.concurrent package. It internally forwards the call to either a forkjoin based task support or a thread pool executor one, depending on what the execution context uses.

By default, parallel collections are parameterized with this task support object, so parallel collections share the same execution context backend as the rest of the scala.concurrent package.

This tasks implementation uses execution contexts to spawn a parallel computation.

As an optimization, it internally checks whether the execution context is the standard implementation based on fork/join pools, and if it is, creates a ForkJoinTaskSupport that shares the same pool to forward its request to it.

Otherwise, it uses an execution context exclusive Tasks implementation to divide the tasks into smaller chunks and execute operations on it.

A task support that uses a fork join pool to schedule tasks.

Parallel iterators allow splitting and provide a remaining method to obtain the number of elements remaining in the iterator.

A template trait for parallel iterable collections.

This is a base trait for Scala parallel collections. It defines behaviour common to all parallel collections. Concrete parallel collections should inherit this trait and ParIterable if they want to define specific combiner factories.

Parallel operations are implemented with divide and conquer style algorithms that parallelize well. The basic idea is to split the collection into smaller parts until they are small enough to be operated on sequentially.

All of the parallel operations are implemented as tasks within this trait. Tasks rely on the concept of splitters, which extend iterators. Every parallel collection defines:

def splitter: IterableSplitter[T]

which returns an instance of IterableSplitter[T], which is a subtype of Splitter[T]. Splitters have a method remaining to check the remaining number of elements, and method split which is defined by splitters. Method split divides the splitters iterate over into disjunct subsets:

def split: Seq[Splitter]

which splits the splitter into a sequence of disjunct subsplitters. This is typically a very fast operation which simply creates wrappers around the receiver collection. This can be repeated recursively.

Tasks are scheduled for execution through a scala.collection.parallel.TaskSupport object, which can be changed through the tasksupport setter of the collection.

Method newCombiner produces a new combiner. Combiners are an extension of builders. They provide a method combine which combines two combiners and returns a combiner containing elements of both combiners. This method can be implemented by aggressively copying all the elements into the new combiner or by lazily binding their results. It is recommended to avoid copying all of the elements for performance reasons, although that cost might be negligible depending on the use case. Standard parallel collection combiners avoid copying when merging results, relying either on a two-step lazy construction or specific data-structure properties.

Methods:

def seq: Sequential
def par: Repr

produce the sequential or parallel implementation of the collection, respectively. Method par just returns a reference to this parallel collection. Method seq is efficient - it will not copy the elements. Instead, it will create a sequential version of the collection using the same underlying data structure. Note that this is not the case for sequential collections in general - they may copy the elements and produce a different underlying data structure.

The combination of methods toMap, toSeq or toSet along with par and seq is a flexible way to change between different collection types.

Since this trait extends the GenIterable trait, methods like size must also be implemented in concrete collections, while iterator forwards to splitter by default.

Each parallel collection is bound to a specific fork/join pool, on which dormant worker threads are kept. The fork/join pool contains other information such as the parallelism level, that is, the number of processors used. When a collection is created, it is assigned the default fork/join pool found in the scala.parallel package object.

Parallel collections are not necessarily ordered in terms of the foreach operation (see Traversable). Parallel sequences have a well defined order for iterators - creating an iterator and traversing the elements linearly will always yield the same order. However, bulk operations such as foreach, map or filter always occur in undefined orders for all parallel collections.

Existing parallel collection implementations provide strict parallel iterators. Strict parallel iterators are aware of the number of elements they have yet to traverse. It's also possible to provide non-strict parallel iterators, which do not know the number of elements remaining. To do this, the new collection implementation must override isStrictSplitterCollection to false. This will make some operations unavailable.

To create a new parallel collection, extend the ParIterable trait, and implement size, splitter, newCombiner and seq. Having an implicit combiner factory requires extending this trait in addition, as well as providing a companion object, as with regular collections.

Method size is implemented as a constant time operation for parallel collections, and parallel collection operations rely on this assumption.

The higher-order functions passed to certain operations may contain side-effects. Since implementations of bulk operations may not be sequential, this means that side-effects may not be predictable and may produce data-races, deadlocks or invalidation of state if care is not taken. It is up to the programmer to either avoid using side-effects or to use some form of synchronization when accessing mutable data.

A template trait for parallel collections of type ParIterable[T].

This is a base trait for Scala parallel collections. It defines behaviour common to all parallel collections. Concrete parallel collections should inherit this trait and ParIterable if they want to define specific combiner factories.

Parallel operations are implemented with divide and conquer style algorithms that parallelize well. The basic idea is to split the collection into smaller parts until they are small enough to be operated on sequentially.

All of the parallel operations are implemented as tasks within this trait. Tasks rely on the concept of splitters, which extend iterators. Every parallel collection defines:

def splitter: IterableSplitter[T]

which returns an instance of IterableSplitter[T], which is a subtype of Splitter[T]. Splitters have a method remaining to check the remaining number of elements, and method split which is defined by splitters. Method split divides the splitters iterate over into disjunct subsets:

def split: Seq[Splitter]

which splits the splitter into a sequence of disjunct subsplitters. This is typically a very fast operation which simply creates wrappers around the receiver collection. This can be repeated recursively.

Tasks are scheduled for execution through a scala.collection.parallel.TaskSupport object, which can be changed through the tasksupport setter of the collection.

Method newCombiner produces a new combiner. Combiners are an extension of builders. They provide a method combine which combines two combiners and returns a combiner containing elements of both combiners. This method can be implemented by aggressively copying all the elements into the new combiner or by lazily binding their results. It is recommended to avoid copying all of the elements for performance reasons, although that cost might be negligible depending on the use case. Standard parallel collection combiners avoid copying when merging results, relying either on a two-step lazy construction or specific data-structure properties.

Methods:

def seq: Sequential
def par: Repr

produce the sequential or parallel implementation of the collection, respectively. Method par just returns a reference to this parallel collection. Method seq is efficient - it will not copy the elements. Instead, it will create a sequential version of the collection using the same underlying data structure. Note that this is not the case for sequential collections in general - they may copy the elements and produce a different underlying data structure.

The combination of methods toMap, toSeq or toSet along with par and seq is a flexible way to change between different collection types.

Since this trait extends the GenIterable trait, methods like size must also be implemented in concrete collections, while iterator forwards to splitter by default.

Each parallel collection is bound to a specific fork/join pool, on which dormant worker threads are kept. The fork/join pool contains other information such as the parallelism level, that is, the number of processors used. When a collection is created, it is assigned the default fork/join pool found in the scala.parallel package object.

Parallel collections are not necessarily ordered in terms of the foreach operation (see Traversable). Parallel sequences have a well defined order for iterators - creating an iterator and traversing the elements linearly will always yield the same order. However, bulk operations such as foreach, map or filter always occur in undefined orders for all parallel collections.

Existing parallel collection implementations provide strict parallel iterators. Strict parallel iterators are aware of the number of elements they have yet to traverse. It's also possible to provide non-strict parallel iterators, which do not know the number of elements remaining. To do this, the new collection implementation must override isStrictSplitterCollection to false. This will make some operations unavailable.

To create a new parallel collection, extend the ParIterable trait, and implement size, splitter, newCombiner and seq. Having an implicit combiner factory requires extending this trait in addition, as well as providing a companion object, as with regular collections.

Method size is implemented as a constant time operation for parallel collections, and parallel collection operations rely on this assumption.

The higher-order functions passed to certain operations may contain side-effects. Since implementations of bulk operations may not be sequential, this means that side-effects may not be predictable and may produce data-races, deadlocks or invalidation of state if care is not taken. It is up to the programmer to either avoid using side-effects or to use some form of synchronization when accessing mutable data.

A template trait for parallel maps.

The higher-order functions passed to certain operations may contain side-effects. Since implementations of bulk operations may not be sequential, this means that side-effects may not be predictable and may produce data-races, deadlocks or invalidation of state if care is not taken. It is up to the programmer to either avoid using side-effects or to use some form of synchronization when accessing mutable data.

A template trait for mutable parallel maps. This trait is to be mixed in with concrete parallel maps to override the representation type.

The higher-order functions passed to certain operations may contain side-effects. Since implementations of bulk operations may not be sequential, this means that side-effects may not be predictable and may produce data-races, deadlocks or invalidation of state if care is not taken. It is up to the programmer to either avoid using side-effects or to use some form of synchronization when accessing mutable data.

A template trait for parallel sequences.

Parallel sequences inherit the Seq trait. Their indexing and length computations are defined to be efficient. Like their sequential counterparts they always have a defined order of elements. This means they will produce resulting parallel sequences in the same way sequential sequences do. However, the order in which they perform bulk operations on elements to produce results is not defined and is generally nondeterministic. If the higher-order functions given to them produce no sideeffects, then this won't be noticeable.

This trait defines a new, more general split operation and reimplements the split operation of ParallelIterable trait using the new split operation.

The higher-order functions passed to certain operations may contain side-effects. Since implementations of bulk operations may not be sequential, this means that side-effects may not be predictable and may produce data-races, deadlocks or invalidation of state if care is not taken. It is up to the programmer to either avoid using side-effects or to use some form of synchronization when accessing mutable data.

A template trait for sequences of type ParSeq[T], representing parallel sequences with element type T.

Parallel sequences inherit the Seq trait. Their indexing and length computations are defined to be efficient. Like their sequential counterparts they always have a defined order of elements. This means they will produce resulting parallel sequences in the same way sequential sequences do. However, the order in which they perform bulk operations on elements to produce results is not defined and is generally nondeterministic. If the higher-order functions given to them produce no sideeffects, then this won't be noticeable.

This trait defines a new, more general split operation and reimplements the split operation of ParallelIterable trait using the new split operation.

A template trait for parallel sets.

The higher-order functions passed to certain operations may contain side-effects. Since implementations of bulk operations may not be sequential, this means that side-effects may not be predictable and may produce data-races, deadlocks or invalidation of state if care is not taken. It is up to the programmer to either avoid using side-effects or to use some form of synchronization when accessing mutable data.

A template trait for parallel sets. This trait is mixed in with concrete parallel sets to override the representation type.

The higher-order functions passed to certain operations may contain side-effects. Since implementations of bulk operations may not be sequential, this means that side-effects may not be predictable and may produce data-races, deadlocks or invalidation of state if care is not taken. It is up to the programmer to either avoid using side-effects or to use some form of synchronization when accessing mutable data.

A precise splitter (or a precise split iterator) can be split into arbitrary number of splitters that traverse disjoint subsets of arbitrary sizes.

Implementors might want to override the parameterless split method for efficiency.

Parallel sequence iterators allow splitting into arbitrary subsets.

A splitter (or a split iterator) can be split into more splitters that traverse over disjoint subsets of elements.

A trait implementing the scheduling of a parallel collection operation.

Parallel collections are modular in the way operations are scheduled. Each parallel collection is parameterized with a task support object which is responsible for scheduling and load-balancing tasks to processors.

A task support object can be changed in a parallel collection after it has been created, but only during a quiescent period, i.e. while there are no concurrent invocations to parallel collection methods.

There are currently a few task support implementations available for parallel collections. The scala.collection.parallel.ForkJoinTaskSupport uses a fork-join pool internally.

The scala.collection.parallel.ExecutionContextTaskSupport uses the default execution context implementation found in scala.concurrent, and it reuses the thread pool used in scala.concurrent.

The execution context task support is set to each parallel collection by default, so parallel collections reuse the same fork-join pool as the future API.

Here is a way to change the task support of a parallel collection:

import scala.collection.parallel._
val pc = mutable.ParArray(1, 2, 3)
pc.tasksupport = new ForkJoinTaskSupport(
  new java.util.concurrent.ForkJoinPool(2))

Composite throwable - thrown when multiple exceptions are thrown at the same time.

A task support that uses a thread pool executor to schedule tasks.

An implementation of tasks objects based on the Java thread pooling API.