jmmcg::ppd::pool_traits::private_::thread_pool_queue_details_front_batch_priority_queue< pool_traits::work_distribution_mode_t::queue_model_t::pool_owns_queue, EvSts, Stats, V, Comp, GSSk > Struct Template Reference

#include <thread_pool_aspects.hpp>

Collaboration diagram for jmmcg::ppd::pool_traits::private_::thread_pool_queue_details_front_batch_priority_queue< pool_traits::work_distribution_mode_t::queue_model_t::pool_owns_queue, EvSts, Stats, V, Comp, GSSk >:

Public Types
using	queue_model = pool_traits::work_distribution_mode_t::queue_model_t::pool_owns_queue
using	element_type = shared_ptr< V, typename V::base_t::lock_traits >
using	exit_requested_type = typename lock::new_event_signal< EvSts, EvSts::unsignalled, EvSts::exit_requested, typename element_type::lock_traits::critical_section_type >
	To be used by the thread_pool to signal that it requires the pool_threads it manages should exit, or have closure_base-derived closure to process. More...
using	lock_all_type = typename lock::any_order::all< exit_requested_type::lock_traits::api_type, typename exit_requested_type::lock_traits::model_type, typename exit_requested_type::locker_type, typename exit_requested_type::locker_type >
using	key_compare = typename ppd::private_::def_key_compare< Comp, element_type >::result
using	container_type = front_batch< priority_queue< std::priority_queue< element_type, std::vector< element_type >, key_compare >, typename exit_requested_type::locker_type, typename exit_requested_type::locker_type::write_lock_type, signalling< exit_requested_type >, std::array< element_type, GSSk > >, GSSk >
	The signalled_work_queue_type is an adapted list to add some thread-safety, because we need to perform multiple operations on it, atomically. More...
using	statistics_type = Stats< typename container_type::size_type >
using	have_work_type = typename container_type::have_work_type

Public Member Functions
	thread_pool_queue_details_front_batch_priority_queue () FORCE_INLINE

Public Attributes
exit_requested_type	exit_requested_
	This is one exit event for all of the threads in the pool: so that they can exit in parallel. More...
container_type	signalled_work_queue

Static Public Attributes
static constexpr generic_traits::memory_access_modes	memory_access_mode =container_type::memory_access_mode

Detailed Description

template<class EvSts, template< class > class Stats, class V, template< class > class Comp, unsigned long GSSk>
struct jmmcg::ppd::pool_traits::private_::thread_pool_queue_details_front_batch_priority_queue< pool_traits::work_distribution_mode_t::queue_model_t::pool_owns_queue, EvSts, Stats, V, Comp, GSSk >

Definition at line 315 of file thread_pool_aspects.hpp.

Member Typedef Documentation

container_type

template<class EvSts , template< class > class Stats, class V , template< class > class Comp, unsigned long GSSk>

using jmmcg::ppd::pool_traits::private_::thread_pool_queue_details_front_batch_priority_queue< pool_traits::work_distribution_mode_t::queue_model_t::pool_owns_queue, EvSts, Stats, V, Comp, GSSk >::container_type = front_batch< priority_queue< std::priority_queue<element_type, std::vector<element_type>, key_compare>, typename exit_requested_type::locker_type, typename exit_requested_type::locker_type::write_lock_type, signalling<exit_requested_type>, std::array<element_type, GSSk> >, GSSk >

The signalled_work_queue_type is an adapted list to add some thread-safety, because we need to perform multiple operations on it, atomically.

Basically this has been defined as a FIFO queue that has locking semantics defined by the specific collection used, and implements a simple list-based scheduling algorithm that assumes that each work-item is the same work-load.

Todo:

JMG: if the work-load has a user-supplied estimate of the time it would take to complete, then we could implement a more sophisticated and better algorithm, e.g. in [1].

The algorithmic complexity of the operations on this container is based upon the complexity of the operations on the underlying container, in this case an instrusive::slist. This specifically implies that adding and removing an item takes constant time.

This queue would simply implement a bakers' scheduling algorithm: i.e. the active pool_thread would remove k items from the front of the queue to be mutated by that thread.

1. The items in the queue would be a list of work items. i.e. V should be a container.
2. When an item is added to the queue, it would be appended to the last item's queue, leaving some distance between the last & first item, e.g. the number of threads in the pool, to try and load-balance the work-load.
3. The pool_threads would loop through the list of items that they extract to perform the mutations.
4. The items in the queue would be a list of work items.
5. The pool_threads would remove at most GSSk work-items from the queue and loop through the list of items that it extracted to perform those mutations.

See also

back_batch

queue, funky_queue

intrusive::slist

[1] Casanova, H., Legrand, A., Robert, Y., "Parallel Algorithms", CRC Press, 2008.

Definition at line 373 of file thread_pool_aspects.hpp.

element_type

template<class EvSts , template< class > class Stats, class V , template< class > class Comp, unsigned long GSSk>

using jmmcg::ppd::pool_traits::private_::thread_pool_queue_details_front_batch_priority_queue< pool_traits::work_distribution_mode_t::queue_model_t::pool_owns_queue, EvSts, Stats, V, Comp, GSSk >::element_type = shared_ptr< V, typename V::base_t::lock_traits >

We don't actually need this to be guaranteed lockfree, as locking is done elsewhere, so we can gain a smidge of performance by using raw pointers.

Definition at line 323 of file thread_pool_aspects.hpp.

exit_requested_type

template<class EvSts , template< class > class Stats, class V , template< class > class Comp, unsigned long GSSk>

using jmmcg::ppd::pool_traits::private_::thread_pool_queue_details_front_batch_priority_queue< pool_traits::work_distribution_mode_t::queue_model_t::pool_owns_queue, EvSts, Stats, V, Comp, GSSk >::exit_requested_type = typename lock::new_event_signal< EvSts, EvSts::unsignalled, EvSts::exit_requested, typename element_type::lock_traits::critical_section_type >

To be used by the thread_pool to signal that it requires the pool_threads it manages should exit, or have closure_base-derived closure to process.

This is modelling the functionality of "WaitForMultipleObjects(...)" in the Win32 API.

Definition at line 333 of file thread_pool_aspects.hpp.

have_work_type

template<class EvSts , template< class > class Stats, class V , template< class > class Comp, unsigned long GSSk>

using jmmcg::ppd::pool_traits::private_::thread_pool_queue_details_front_batch_priority_queue< pool_traits::work_distribution_mode_t::queue_model_t::pool_owns_queue, EvSts, Stats, V, Comp, GSSk >::have_work_type = typename container_type::have_work_type

Definition at line 375 of file thread_pool_aspects.hpp.

key_compare

template<class EvSts , template< class > class Stats, class V , template< class > class Comp, unsigned long GSSk>

using jmmcg::ppd::pool_traits::private_::thread_pool_queue_details_front_batch_priority_queue< pool_traits::work_distribution_mode_t::queue_model_t::pool_owns_queue, EvSts, Stats, V, Comp, GSSk >::key_compare = typename ppd::private_::def_key_compare<Comp, element_type>::result

Definition at line 340 of file thread_pool_aspects.hpp.

lock_all_type

template<class EvSts , template< class > class Stats, class V , template< class > class Comp, unsigned long GSSk>

using jmmcg::ppd::pool_traits::private_::thread_pool_queue_details_front_batch_priority_queue< pool_traits::work_distribution_mode_t::queue_model_t::pool_owns_queue, EvSts, Stats, V, Comp, GSSk >::lock_all_type = typename lock::any_order::all< exit_requested_type::lock_traits::api_type, typename exit_requested_type::lock_traits::model_type, typename exit_requested_type::locker_type, typename exit_requested_type::locker_type >

Definition at line 339 of file thread_pool_aspects.hpp.

queue_model

template<class EvSts , template< class > class Stats, class V , template< class > class Comp, unsigned long GSSk>

using jmmcg::ppd::pool_traits::private_::thread_pool_queue_details_front_batch_priority_queue< pool_traits::work_distribution_mode_t::queue_model_t::pool_owns_queue, EvSts, Stats, V, Comp, GSSk >::queue_model = pool_traits::work_distribution_mode_t::queue_model_t::pool_owns_queue

Definition at line 316 of file thread_pool_aspects.hpp.

statistics_type

template<class EvSts , template< class > class Stats, class V , template< class > class Comp, unsigned long GSSk>

using jmmcg::ppd::pool_traits::private_::thread_pool_queue_details_front_batch_priority_queue< pool_traits::work_distribution_mode_t::queue_model_t::pool_owns_queue, EvSts, Stats, V, Comp, GSSk >::statistics_type = Stats<typename container_type::size_type>

Definition at line 374 of file thread_pool_aspects.hpp.

Constructor & Destructor Documentation

thread_pool_queue_details_front_batch_priority_queue()

template<class EvSts , template< class > class Stats, class V , template< class > class Comp, unsigned long GSSk>

jmmcg::ppd::pool_traits::private_::thread_pool_queue_details_front_batch_priority_queue< pool_traits::work_distribution_mode_t::queue_model_t::pool_owns_queue, EvSts, Stats, V, Comp, GSSk >::thread_pool_queue_details_front_batch_priority_queue ( )

inline

Definition at line 389 of file thread_pool_aspects.hpp.

References memory_access_mode.

Member Data Documentation

exit_requested_

template<class EvSts , template< class > class Stats, class V , template< class > class Comp, unsigned long GSSk>

exit_requested_type jmmcg::ppd::pool_traits::private_::thread_pool_queue_details_front_batch_priority_queue< pool_traits::work_distribution_mode_t::queue_model_t::pool_owns_queue, EvSts, Stats, V, Comp, GSSk >::exit_requested_

mutable

This is one exit event for all of the threads in the pool: so that they can exit in parallel.

Definition at line 383 of file thread_pool_aspects.hpp.

memory_access_mode

template<class EvSts , template< class > class Stats, class V , template< class > class Comp, unsigned long GSSk>

constexpr generic_traits::memory_access_modes jmmcg::ppd::pool_traits::private_::thread_pool_queue_details_front_batch_priority_queue< pool_traits::work_distribution_mode_t::queue_model_t::pool_owns_queue, EvSts, Stats, V, Comp, GSSk >::memory_access_mode =container_type::memory_access_mode

staticconstexpr

To assist in allowing compile-time computation of the algorithmic order of the threading model.

Definition at line 380 of file thread_pool_aspects.hpp.

Referenced by jmmcg::ppd::lock::new_event_signal< St, UnSig, Pri, api_lock_traits< platform_api, sequential_mode >::critical_section_type >::locker(), and thread_pool_queue_details_front_batch_priority_queue().

signalled_work_queue

template<class EvSts , template< class > class Stats, class V , template< class > class Comp, unsigned long GSSk>

container_type jmmcg::ppd::pool_traits::private_::thread_pool_queue_details_front_batch_priority_queue< pool_traits::work_distribution_mode_t::queue_model_t::pool_owns_queue, EvSts, Stats, V, Comp, GSSk >::signalled_work_queue

This implies that the thread_pool assumes a flat memory model, shared across all processors. This may not reflect reality well, in the face of caches.

Definition at line 387 of file thread_pool_aspects.hpp.

---

The documentation for this struct was generated from the following file:

• core/thread_pool_aspects.hpp