A thread pool is a set of OCaml threads used to do work, where each piece of work is simply a thunk. One creates a thread pool, and then uses add_work to submit work to it. Work is done first-come-first-served by available threads in the pool. Any of the available threads in the pool could be used to do work submitted to the pool (except helper threads, see below).
A thread pool starts with no threads. As work is added, the thread pool creates new threads to do the work, up to the maximum number of allowed threads, max_num_threads, supplied to create. Thread-pool threads never die. They just get created up until max_num_threads is reached and then live forever, doing work. Each thread in the pool is in a loop, waiting for a piece of work, running the thunk, and then repeating. It may be that all the threads in the pool are not doing anything, but in this case, the threads still exist, and are simply blocked waiting for work.
Sometimes one wants work to run in a dedicated thread, e.g. some C libraries require this. To do this, use Helper_thread, see below.
All of the functions exposed by this module are thread safe; they synchronize using a mutex on the thread pool.
One must not call thread-pool functions from a GC finalizer, since a finalizer could run within a thread running a thread-pool function, which already holds the lock, and would therefore deadlock or error when attempting to re-acquire it. This is accomplished elsewhere by using Async finalizers, which are run from ordinary Async jobs, and thus do not hold the thread-pool lock.
One can control the priority and affinity of threads in the pool (priority in the sense of Linux_ext.setpriority). Work added to the pool can optionally be given a priority, and the pool will set the priority of the thread that runs it for the duration of the work. Helper threads can also be given a priority, which will be used for all work run by the helper thread, unless the work has an overriding priority. The thread pool has a "default" priority that will be used for all work and helper threads that have no specified priority. The default priority is the priority in effect when create is called.
Affinity, on the other hand, can only be specified when you create a pool. The default affinity is the affinity in effect when a new thread happens to be created (e.g. when you call add_work).
Behavior is unspecified if work calls setpriority or setaffinity directly.
create ?cpuset ~max_num_threads returns a new thread pool. It is an error if max_num_threads < 1.
If cpuset is specified, then every thread will be affinitized to those CPU cores upon creation.
If cpuset is not specified, then every thread will inherit the affinitization of the thread (typically the main thread) that created it.
cpu_affinity t returns the CPU affinity that t was created with. All threads created by t will be created with this affinity.
val finished_with : t -> unitfinished_with t makes it an error to subsequently call add_work* t or create_helper_thread t. And, once all current work in t is finished, destroys all the threads in t. It is OK to call finished_with multiple times on the same t; subsequent calls will have no effect.
val block_until_finished : t -> unitblock_until_finished t blocks the current thread until thread pool t has finished. One must previously have called finished_with to cause t to start finishing.
val max_num_threads : t -> intmax_num_threads t returns the maximum number of threads that t is allowed to create.
val num_threads : t -> intnum_threads t returns the number of threads that the pool t has created.
val unfinished_work : t -> intunfinished_work t returns the number of jobs that have been submitted to t but haven't yet finished.
default_priority t returns the priority that will be used for work performed by t, unless that work is added with an overriding priority.
add_work ?priority ?name t f enqueues f to be done by some thread in the pool.
Exceptions raised by f are silently ignored.
While the work is run, the name of the thread running the work will be set (via Linux_ext.pr_set_name) to name and the priority of the thread will be set to priority.
It is an error to call add_work t after finished_with t.
val num_work_completed : t -> intval has_unstarted_work : t -> boolhas_unstarted_work t returns true if t has work that it hasn't been assigned to start running in a thread.
A helper thread is a thread with its own dedicated work queue. Work added for the helper thread is guaranteed to be run by that thread. The helper thread only runs work explicitly supplied to it. Helper threads count towards a thread pool's max_num_threads.
create_helper_thread ?priority ?name t takes an available thread from the thread pool and makes it a helper thread, raising if no threads are available or if finished_with t was previously called. The new helper thread runs work with name and priority, except for work that is added with an overriding priority or name. The thread remains a helper thread until finished_with_helper_thread is called, if ever.
become_helper_thread ?priority ?name t should be run from within work supplied to add_work. When become_helper_thread runs, it transitions the current thread into a helper thread.
Other than that, become_helper_thread is like create_helper_thread, except it cannot fail because no threads are available.
add_work_for_helper_thread ?priority ?name t helper_thread f enqueues f on helper_thread's work queue.
Exceptions raised by f are silently ignored.
It is an error to call add_work_for_helper_thread t after finished_with_helper_thread t.
When the helper thread runs f, it will be at the helper thread's name and priority, unless overriden by name or priority.
finished_with_helper_thread t helper_thread informs thread pool t that no future work will be added for helper_thread, and makes it an error to in the future add work for helper_thread. Furthermore, once helper_thread finishes with its last piece of work, it will revert to a general thread-pool thread. It is OK to call finished_with_helper_thread multiple times on the same helper_thread; subsequent calls will have no effect.