Writing single-node HPX applications
Contents
Writing single-node HPX applications#
HPX is a C++ Standard Library for Concurrency and Parallelism. This means that it implements all of the corresponding facilities as defined by the C++ Standard. Additionally, HPX implements functionalities proposed as part of the ongoing C++ standardization process. This section focuses on the features available in HPX for parallel and concurrent computation on a single node, although many of the features presented here are also implemented to work in the distributed case.
Using LCOs#
Lightweight Control Objects (LCOs) provide synchronization for HPX applications. Most of them are familiar from other frameworks, but a few of them work in slightly different ways adapted to HPX. The following synchronization objects are available in HPX:
futurequeueobject_semaphorebarrier
Channels#
Channels combine communication (the exchange of a value) with synchronization (guaranteeing that two calculations (tasks) are in a known state). A channel can transport any number of values of a given type from a sender to a receiver:
hpx::lcos::local::channel<int> c;
hpx::future<int> f = c.get();
HPX_ASSERT(!f.is_ready());
c.set(42);
HPX_ASSERT(f.is_ready());
std::cout << f.get() << std::endl;
Channels can be handed to another thread (or in case of channel components, to other localities), thus establishing a communication channel between two independent places in the program:
void do_something(hpx::lcos::local::receive_channel<int> c,
hpx::lcos::local::send_channel<> done)
{
// prints 43
std::cout << c.get(hpx::launch::sync) << std::endl;
// signal back
done.set();
}
void send_receive_channel()
{
hpx::lcos::local::channel<int> c;
hpx::lcos::local::channel<> done;
hpx::apply(&do_something, c, done);
// send some value
c.set(43);
// wait for thread to be done
done.get().wait();
}
Note how hpx::lcos::local::channel::get without any arguments
returns a future which is ready when a value has been set on the channel. The
launch policy hpx::launch::sync can be used to make
hpx::lcos::local::channel::get block until a value is set and
return the value directly.
A channel component is created on one locality and can be sent to another locality using an action. This example also demonstrates how a channel can be used as a range of values:
// channel components need to be registered for each used type (not needed
// for hpx::lcos::local::channel)
HPX_REGISTER_CHANNEL(double)
void channel_sender(hpx::lcos::channel<double> c)
{
for (double d : c)
hpx::cout << d << std::endl;
}
HPX_PLAIN_ACTION(channel_sender)
void channel()
{
// create the channel on this locality
hpx::lcos::channel<double> c(hpx::find_here());
// pass the channel to a (possibly remote invoked) action
hpx::apply(channel_sender_action(), hpx::find_here(), c);
// send some values to the receiver
std::vector<double> v = {1.2, 3.4, 5.0};
for (double d : v)
c.set(d);
// explicitly close the communication channel (implicit at destruction)
c.close();
}
Composable guards#
Composable guards operate in a manner similar to locks, but are applied only to asynchronous functions. The guard (or guards) is automatically locked at the beginning of a specified task and automatically unlocked at the end. Because guards are never added to an existing task’s execution context, the calling of guards is freely composable and can never deadlock.
To call an application with a single guard, simply declare the guard and call run_guarded() with a function (task):
hpx::lcos::local::guard gu;
run_guarded(gu,task);
If a single method needs to run with multiple guards, use a guard set:
boost::shared<hpx::lcos::local::guard> gu1(new hpx::lcos::local::guard());
boost::shared<hpx::lcos::local::guard> gu2(new hpx::lcos::local::guard());
gs.add(*gu1);
gs.add(*gu2);
run_guarded(gs,task);
Guards use two atomic operations (which are not called repeatedly) to manage what they do, so overhead should be extremely low. The following guards are available in HPX:
conditional_triggercounting_semaphoredatafloweventmutexoncerecursive_mutexspinlockspinlock_no_backofftrigger
Extended facilities for futures#
Concurrency is about both decomposing and composing the program from the parts that work well individually and together. It is in the composition of connected and multicore components where today’s C++ libraries are still lacking.
The functionality of std::future offers a partial solution. It allows for
the separation of the initiation of an operation and the act of waiting for its
result; however, the act of waiting is synchronous. In communication-intensive
code this act of waiting can be unpredictable, inefficient and simply
frustrating. The example below illustrates a possible synchronous wait using
futures:
#include <future>
using namespace std;
int main()
{
future<int> f = async([]() { return 123; });
int result = f.get(); // might block
}
For this reason, HPX implements a set of extensions to std::future (as
proposed by __cpp11_n4107__). This proposal introduces the following key
asynchronous operations to hpx::future, hpx::shared_future and
hpx::async, which enhance and enrich these facilities.
Facility |
Description |
|
In asynchronous programming, it is very common for one asynchronous
operation, on completion, to invoke a second operation and pass data to
it. The current C++ standard does not allow one to register a
continuation to a future. With |
unwrapping constructor for |
In some scenarios, you might want to create a future that returns another future, resulting in nested futures. Although it is possible to write code to unwrap the outer future and retrieve the nested future and its result, such code is not easy to write because users must handle exceptions and it may cause a blocking call. Unwrapping can allow users to mitigate this problem by doing an asynchronous call to unwrap the outermost future. |
|
There are often situations where a |
|
Some functions may know the value at the point of construction. In these
cases the value is immediately available, but needs to be returned as a
future. By using |
The standard also omits the ability to compose multiple futures. This is a common pattern that is ubiquitous in other asynchronous frameworks and is absolutely necessary in order to make C++ a powerful asynchronous programming language. Not including these functions is synonymous to Boolean algebra without AND/OR.
In addition to the extensions proposed by N4313, HPX adds functions allowing users to compose several futures in a more flexible way.
Facility |
Description |
Comment |
Asynchronously wait for at least one of multiple future or shared_future objects to finish. |
N4313, |
|
Synchronously wait for at least one of multiple future or shared_future objects to finish. |
HPX only |
|
Asynchronously wait for all future and shared_future objects to finish. |
N4313, |
|
Synchronously wait for all future and shared_future objects to finish. |
HPX only |
|
Asynchronously wait for multiple future and shared_future objects to finish. |
HPX only |
|
Synchronously wait for multiple future and shared_future objects to finish. |
HPX only |
|
Asynchronously wait for multiple future and shared_future objects to finish and call a function for each of the future objects as soon as it becomes ready. |
HPX only |
|
Synchronously wait for multiple future and shared_future objects to finish and call a function for each of the future objects as soon as it becomes ready. |
HPX only |
High level parallel facilities#
In preparation for the upcoming C++ Standards, there are currently several proposals targeting different facilities supporting parallel programming. HPX implements (and extends) some of those proposals. This is well aligned with our strategy to align the APIs exposed from HPX with current and future C++ Standards.
At this point, HPX implements several of the C++ Standardization working papers, most notably N4409 (Working Draft, Technical Specification for C++ Extensions for Parallelism), N4411 (Task Blocks), and N4406 (Parallel Algorithms Need Executors).
Using parallel algorithms#
A parallel algorithm is a function template described by this document
which is declared in the (inline) namespace hpx::parallel::v1.
Note
For compilers that do not support inline namespaces, all of the namespace
v1 is imported into the namespace hpx::parallel. The effect is similar
to what inline namespaces would do, namely all names defined in
hpx::parallel::v1 are accessible from the namespace hpx::parallel as
well.
All parallel algorithms are very similar in semantics to their sequential
counterparts (as defined in the namespace std) with an additional formal
template parameter named ExecutionPolicy. The execution policy is generally
passed as the first argument to any of the parallel algorithms and describes the
manner in which the execution of these algorithms may be parallelized and the
manner in which they apply user-provided function objects.
The applications of function objects in parallel algorithms invoked with an
execution policy object of type hpx::execution::sequenced_policy or
hpx::execution::sequenced_task_policy execute in sequential order. For
hpx::execution::sequenced_policy the execution happens in the calling thread.
The applications of function objects in parallel algorithms invoked with an
execution policy object of type hpx::execution::parallel_policy or
hpx::execution::parallel_task_policy are permitted to execute in an unordered
fashion in unspecified threads, and are indeterminately sequenced within each
thread.
Important
It is the caller’s responsibility to ensure correctness, such as making sure that the invocation does not introduce data races or deadlocks.
The applications of function objects in parallel algorithms invoked with an
execution policy of type hpx::execution::parallel_unsequenced_policy is, in HPX,
equivalent to the use of the execution policy hpx::execution::parallel_policy.
Algorithms invoked with an execution policy object of type hpx::parallel::v1::execution_policy
execute internally as if invoked with the contained execution policy object. No
exception is thrown when an hpx::parallel::v1::execution_policy contains an execution policy of
type hpx::execution::sequenced_task_policy or hpx::execution::parallel_task_policy
(which normally turn the algorithm into its asynchronous version). In this case
the execution is semantically equivalent to the case of passing a
hpx::execution::sequenced_policy or hpx::execution::parallel_policy contained in the
hpx::parallel::v1::execution_policy object respectively.
Parallel exceptions#
During the execution of a standard parallel algorithm, if temporary memory
resources are required by any of the algorithms and no memory is available, the
algorithm throws a std::bad_alloc exception.
During the execution of any of the parallel algorithms, if the application of a function object terminates with an uncaught exception, the behavior of the program is determined by the type of execution policy used to invoke the algorithm:
If the execution policy object is of type
hpx::execution::parallel_unsequenced_policy,hpx::terminateshall be called.If the execution policy object is of type
hpx::execution::sequenced_policy,hpx::execution::sequenced_task_policy,hpx::execution::parallel_policy, orhpx::execution::parallel_task_policy, the execution of the algorithm terminates with anhpx::exception_listexception. All uncaught exceptions thrown during the application of user-provided function objects shall be contained in thehpx::exception_list.
For example, the number of invocations of the user-provided function object in
for_each is unspecified. When hpx::parallel::v1::for_each is executed sequentially, only one
exception will be contained in the hpx::exception_list object.
These guarantees imply that, unless the algorithm has failed to allocate memory
and terminated with std::bad_alloc, all exceptions thrown during the
execution of the algorithm are communicated to the caller. It is unspecified
whether an algorithm implementation will “forge ahead” after encountering and
capturing a user exception.
The algorithm may terminate with the std::bad_alloc exception even if one or
more user-provided function objects have terminated with an exception. For
example, this can happen when an algorithm fails to allocate memory while
creating or adding elements to the hpx::exception_list object.
Parallel algorithms#
HPX provides implementations of the following parallel algorithms:
Name |
Description |
In header |
Algorithm page at cppreference.com |
Computes the differences between adjacent elements in a range. |
|
||
Checks if a predicate is |
|
||
Checks if a predicate is |
|
||
Returns the number of elements equal to a given value. |
|
||
Returns the number of elements satisfying a specific criteria. |
|
||
Determines if two sets of elements are the same. |
|
||
Finds the first element equal to a given value. |
|
||
Finds the last sequence of elements in a certain range. |
|
||
Searches for any one of a set of elements. |
|
||
Finds the first element satisfying a specific criteria. |
|
||
Finds the first element not satisfying a specific criteria. |
|
||
Applies a function to a range of elements. |
|
||
Applies a function to a number of elements. |
|
||
Checks if a range of values is lexicographically less than another range of values. |
|
||
|
Finds the first position where two ranges differ. |
|
|
Checks if a predicate is |
|
||
Searches for a range of elements. |
|
||
Searches for a number consecutive copies of an element in a range. |
|
Name |
Description |
In header |
Algorithm page at cppreference.com |
Copies a range of elements to a new location. |
|
||
Copies a number of elements to a new location. |
|
||
Copies the elements from a range to a new location for which the given predicate is |
|
||
Moves a range of elements to a new location. |
|
||
Assigns a range of elements a certain value. |
|
||
Assigns a value to a number of elements. |
|
||
Saves the result of a function in a range. |
|
||
Saves the result of N applications of a function. |
|
||
Removes the elements from a range that are equal to the given value. |
|
||
Removes the elements from a range that are equal to the given predicate is |
|
||
Copies the elements from a range to a new location that are not equal to the given value. |
|
||
Copies the elements from a range to a new location for which the given predicate is |
|
||
Replaces all values satisfying specific criteria with another value. |
|
||
Replaces all values satisfying specific criteria with another value. |
|
||
Copies a range, replacing elements satisfying specific criteria with another value. |
|
||
Copies a range, replacing elements satisfying specific criteria with another value. |
|
||
Reverses the order elements in a range. |
|
||
Creates a copy of a range that is reversed. |
|
||
|
Rotates the order of elements in a range. |
|
|
|
Copies and rotates a range of elements. |
|
|
Shifts the elements in the range left by n positions. |
|
||
Shifts the elements in the range right by n positions. |
|
||
Swaps two ranges of elements. |
|
||
Applies a function to a range of elements. |
|
||
Eliminates all but the first element from every consecutive group of equivalent elements from a range. |
|
||
Copies the elements from one range to another in such a way that there are no consecutive equal elements. |
|
Name |
Description |
In header |
Algorithm page at cppreference.com |
Merges two sorted ranges. |
|
||
Merges two ordered ranges in-place. |
|
||
Returns true if one set is a subset of another. |
|
||
Computes the difference between two sets. |
|
||
Computes the intersection of two sets. |
|
||
Computes the symmetric difference between two sets. |
|
||
Computes the union of two sets. |
|
Name |
Description |
In header |
Algorithm page at cppreference.com |
Returns |
|
||
Returns the first element that breaks a max heap. |
|
||
Constructs a max heap in the range [first, last). |
|
Name |
Description |
In header |
Algorithm page at cppreference.com |
Returns the largest element in a range. |
|
||
Returns the smallest element in a range. |
|
||
Returns the smallest and the largest element in a range. |
|
Name |
Description |
In header |
Algorithm page at cppreference.com |
Calculates the difference between each element in an input range and the preceding element. |
|
||
Does an exclusive parallel scan over a range of elements. |
|
||
Sums up a range of elements. |
|
||
Does an inclusive parallel scan over a range of elements. |
|
||
|
Performs an inclusive scan on consecutive elements with matching keys,
with a reduction to output only the final sum for each key. The key
sequence |
|
|
Sums up a range of elements after applying a function. Also, accumulates the inner products of two input ranges. |
|
||
Does an inclusive parallel scan over a range of elements after applying a function. |
|
||
|
Does an exclusive parallel scan over a range of elements after applying a function. |
|
Name |
Description |
In header |
Algorithm page at cppreference.com |
Destroys a range of objects. |
|
||
Destroys a range of objects. |
|
||
Copies a range of objects to an uninitialized area of memory. |
|
||
Copies a number of objects to an uninitialized area of memory. |
|
||
Copies a range of objects to an uninitialized area of memory. |
|
||
Copies a number of objects to an uninitialized area of memory. |
|
||
Copies an object to an uninitialized area of memory. |
|
||
Copies an object to an uninitialized area of memory. |
|
||
Moves a range of objects to an uninitialized area of memory. |
|
||
Moves a number of objects to an uninitialized area of memory. |
|
||
Constructs objects in an uninitialized area of memory. |
|
||
Constructs objects in an uninitialized area of memory. |
|
Name |
Description |
In header |
Implements loop functionality over a range specified by integral or iterator bounds. |
|
|
Implements loop functionality over a range specified by integral or iterator bounds. |
|
|
Implements loop functionality over a range specified by integral or iterator bounds. |
|
|
Implements loop functionality over a range specified by integral or iterator bounds. |
|
Executor parameters and executor parameter traits#
HPX introduces the notion of execution parameters and execution parameter traits. At this point, the only parameter that can be customized is the size of the chunks of work executed on a single HPX thread (such as the number of loop iterations combined to run as a single task).
An executor parameter object is responsible for exposing the calculation of the size of the chunks scheduled. It abstracts the (potentially platform-specific) algorithms of determining those chunk sizes.
The way executor parameters are implemented is aligned with the way executors
are implemented. All functionalities of concrete executor parameter types are
exposed and accessible through a corresponding
hpx::parallel::executor_parameter_traits type.
With executor_parameter_traits, clients access all types of executor
parameters uniformly:
std::size_t chunk_size =
executor_parameter_traits<my_parameter_t>::get_chunk_size(my_parameter,
my_executor, [](){ return 0; }, num_tasks);
This call synchronously retrieves the size of a single chunk of loop iterations
(or similar) to combine for execution on a single HPX thread if the overall
number of tasks to schedule is given by num_tasks. The lambda function
exposes a means of test-probing the execution of a single iteration for
performance measurement purposes. The execution parameter type might dynamically
determine the execution time of one or more tasks in order to calculate the
chunk size; see hpx::execution::auto_chunk_size for an example of
this executor parameter type.
Other functions in the interface exist to discover whether an executor parameter
type should be invoked once (i.e., it returns a static chunk size; see
hpx::execution::static_chunk_size) or whether it should be invoked
for each scheduled chunk of work (i.e., it returns a variable chunk size; for an
example, see hpx::execution::guided_chunk_size).
Although this interface appears to require executor parameter type authors to implement all different basic operations, none are required. In practice, all operations have sensible defaults. However, some executor parameter types will naturally specialize all operations for maximum efficiency.
HPX implements the following executor parameter types:
hpx::execution::auto_chunk_size: Loop iterations are divided into pieces and then assigned to threads. The number of loop iterations combined is determined based on measurements of how long the execution of 1% of the overall number of iterations takes. This executor parameter type makes sure that as many loop iterations are combined as necessary to run for the amount of time specified.hpx::execution::static_chunk_size: Loop iterations are divided into pieces of a given size and then assigned to threads. If the size is not specified, the iterations are, if possible, evenly divided contiguously among the threads. This executor parameters type is equivalent to OpenMP’s STATIC scheduling directive.hpx::execution::dynamic_chunk_size: Loop iterations are divided into pieces of a given size and then dynamically scheduled among the cores; when a core finishes one chunk, it is dynamically assigned another. If the size is not specified, the default chunk size is 1. This executor parameter type is equivalent to OpenMP’s DYNAMIC scheduling directive.hpx::execution::guided_chunk_size: Iterations are dynamically assigned to cores in blocks as cores request them until no blocks remain to be assigned. This is similar todynamic_chunk_sizeexcept that the block size decreases each time a number of loop iterations is given to a thread. The size of the initial block is proportional tonumber_of_iterations / number_of_cores. Subsequent blocks are proportional tonumber_of_iterations_remaining / number_of_cores. The optional chunk size parameter defines the minimum block size. The default minimal chunk size is 1. This executor parameter type is equivalent to OpenMP’s GUIDED scheduling directive.
Using task blocks#
The define_task_block, run and the wait functions implemented based
on N4411 are based on the task_block concept that is a part of the
common subset of the Microsoft Parallel Patterns Library (PPL) and the Intel Threading Building Blocks (TBB) libraries.
These implementations adopt a simpler syntax than exposed by those libraries— one that is influenced by language-based concepts, such as spawn and sync from Cilk++ and async and finish from X10. They improve on existing practice in the following ways:
The exception handling model is simplified and more consistent with normal C++ exceptions.
Most violations of strict fork-join parallelism can be enforced at compile time (with compiler assistance, in some cases).
The syntax allows scheduling approaches other than child stealing.
Consider an example of a parallel traversal of a tree, where a user-provided function compute is applied to each node of the tree, returning the sum of the results:
template <typename Func>
int traverse(node& n, Func && compute)
{
int left = 0, right = 0;
define_task_block(
[&](task_block<>& tr) {
if (n.left)
tr.run([&] { left = traverse(*n.left, compute); });
if (n.right)
tr.run([&] { right = traverse(*n.right, compute); });
});
return compute(n) + left + right;
}
The example above demonstrates the use of two of the functions,
hpx::parallel::define_task_block and the
hpx::parallel::task_block::run member function of a
hpx::parallel::task_block.
The task_block function delineates a region in a program code potentially
containing invocations of threads spawned by the run member function of the
task_block class. The run function spawns an HPX thread, a unit of
work that is allowed to execute in parallel with respect to the caller. Any
parallel tasks spawned by run within the task block are joined back to a
single thread of execution at the end of the define_task_block. run
takes a user-provided function object f and starts it asynchronously—i.e.,
it may return before the execution of f completes. The HPX scheduler may
choose to run f immediately or delay running f until compute resources
become available.
A task_block can be constructed only by define_task_block because it has
no public constructors. Thus, run can be invoked directly or indirectly
only from a user-provided function passed to define_task_block:
void g();
void f(task_block<>& tr)
{
tr.run(g); // OK, invoked from within task_block in h
}
void h()
{
define_task_block(f);
}
int main()
{
task_block<> tr; // Error: no public constructor
tr.run(g); // No way to call run outside of a define_task_block
return 0;
}
Extensions for task blocks#
Using execution policies with task blocks#
HPX implements some extensions for task_block beyond the actual
standards proposal N4411. The main addition is that a task_block
can be invoked with an execution policy as its first argument, very similar to
the parallel algorithms.
An execution policy is an object that expresses the requirements on the
ordering of functions invoked as a consequence of the invocation of a
task block. Enabling passing an execution policy to define_task_block
gives the user control over the amount of parallelism employed by the
created task_block. In the following example the use of an explicit
par execution policy makes the user’s intent explicit:
template <typename Func>
int traverse(node *n, Func&& compute)
{
int left = 0, right = 0;
define_task_block(
execution::par, // execution::parallel_policy
[&](task_block<>& tb) {
if (n->left)
tb.run([&] { left = traverse(n->left, compute); });
if (n->right)
tb.run([&] { right = traverse(n->right, compute); });
});
return compute(n) + left + right;
}
This also causes the hpx::parallel::v2::task_block object to be a
template in our implementation. The template argument is the type of the
execution policy used to create the task block. The template argument defaults
to hpx::execution::parallel_policy.
HPX still supports calling hpx::parallel::v2::define_task_block
without an explicit execution policy. In this case the task block will run using
the hpx::execution::parallel_policy.
HPX also adds the ability to access the execution policy that was used to
create a given task_block.
Using executors to run tasks#
Often, users want to be able to not only define an execution policy to use by
default for all spawned tasks inside the task block, but also to
customize the execution context for one of the tasks executed by
task_block::run. Adding an optionally passed executor instance to that
function enables this use case:
template <typename Func>
int traverse(node *n, Func&& compute)
{
int left = 0, right = 0;
define_task_block(
execution::par, // execution::parallel_policy
[&](auto& tb) {
if (n->left)
{
// use explicitly specified executor to run this task
tb.run(my_executor(), [&] { left = traverse(n->left, compute); });
}
if (n->right)
{
// use the executor associated with the par execution policy
tb.run([&] { right = traverse(n->right, compute); });
}
});
return compute(n) + left + right;
}
HPX still supports calling hpx::parallel::v2::task_block::run
without an explicit executor object. In this case the task will be run using the
executor associated with the execution policy that was used to call
hpx::parallel::v2::define_task_block.