Prolog threads can exchange data using dynamic predicates, database records, and other globally shared data. These provide no suitable means to wait for data or a condition as they can only be checked in an expensive polling loop. Message queues provide a means for threads to wait for data or conditions without using the CPU.
Each thread has a message queue attached to it that is identified by the thread. Additional queues are created using message_queue_create/1. Explicitly created queues come in two flavours. When given an alias, they must be destroyed by the user. Anonymous message queues are identified by a blob (see section 12.4.10) and subject to garbage collection.
If more than one thread is waiting for messages on the given queue and at least one of these is waiting with a partially instantiated Term, the waiting threads are all sent a wake-up signal, starting a rush for the available messages in the queue. This behaviour can seriously harm performance with many threads waiting on the same queue as all-but-the-winner perform a useless scan of the queue. If there is only one waiting thread or all waiting threads wait with an unbound variable, an arbitrary thread is restarted to scan the queue.201See the documentation for the POSIX thread functions pthread_cond_signal() v.s. pthread_cond_broadcast() for background information.
deadline
option. If both options are provided, the earlier time is effective.
If Time is 0 or 0.0, thread_send_message/3 examines the queue and sends the message if space is available, but does not suspend if no space is available, failing immediately instead.
If Time < 0, thread_send_message/3 fails immediately without sending the message.
true)
times out every 0.25 seconds and checks for signals. If false,
signals are not checked. If a number is specified, we check for signals
every
Time seconds. Smaller times may be used to improved
responsiveness to signals. Larger times may be used to reduce CPU usage.
Please note that non-unifying messages remain in the queue. After the
following has been executed, thread 1 has the term b(gnu)
in its queue and continues execution using A = gnat.
<thread 1> thread_get_message(a(A)), <thread 2> thread_send_message(Thread_1, b(gnu)), thread_send_message(Thread_1, a(gnat)),
Term may contain attributed variables (see section 8), in which case only terms for which the constraints successfully execute are returned. Handle constraints applies for all predicates that extract terms from message queues. For example, we can get the even numbers from a queue using this code:
get_matching_messages(Queue, Pattern, [H|T]) :-
copy_term(Pattern, H),
thread_get_message(Queue, H, [timeout(0)]),
!,
get_matching_messages(Queue, Pattern, T).
get_matching_messages(_, _, []).
even_numbers(Q, List) :-
freeze(Even, Even mod 2 =:= 0),
get_matching_messages(Q, Even, List).
See also thread_peek_message/1.
message_queue_create(Queue,[]). For
compatibility, calling message_queue_create(+Atom) is
equivalent to
message_queue_create(Queue, [alias(Atom)]). New code should
use
message_queue_create/2
to create a named queue.deadline
option. If both options are provided, the earlier time is effective.
If Time is 0 or 0.0, thread_get_message/3 examines the queue but does not suspend if no matching term is available. Note that unlike thread_peek_message/2, a matching term is removed from the queue.
If Time < 0, thread_get_message/3 fails immediately without removing any message from the queue.
true)
times out every 0.25 seconds and checks for signals. If false,
signals are not checked. If a number is specified, we check for signals
every
Time seconds. Smaller times may be used to improved
responsiveness to signals. Larger times may be used to reduce CPU usage.
The size(Size) property is always present and may be
used to enumerate the created message queues. Note that this predicate
does
not enumerate threads, but can be used to query the properties
of the default queue of a thread.
Explicit message queues are designed with the worker-pool model in mind, where multiple threads wait on a single queue and pick up the first goal to execute. Below is a simple implementation where the workers execute arbitrary Prolog goals. Note that this example provides no means to tell when all work is done. This must be realised using additional synchronisation.
%% create_workers(?Id, +N)
%
% Create a pool with Id and number of workers.
% After the pool is created, post_job/1 can be used to
% send jobs to the pool.
create_workers(Id, N) :-
message_queue_create(Id),
forall(between(1, N, _),
thread_create(do_work(Id), _, [])).
do_work(Id) :-
repeat,
thread_get_message(Id, Goal),
( catch(Goal, E, print_message(error, E))
-> true
; print_message(error, goal_failed(Goal, worker(Id)))
),
fail.
%% post_job(+Id, +Goal)
%
% Post a job to be executed by one of the pool's workers.
post_job(Id, Goal) :-
thread_send_message(Id, Goal).
While message queues realizes communicating agents sharing the same program and optionally dynamic data, the predicate thread_wait/2 facilitates agents that communicate based on a shared blackboard. An important difference is were message queues require the sender and receiver to know about the queue used to communicate and every message can wakeup at most one thread, the blackboard model allows any number (including zero) of threads to listen to changes on the blackboard. Any module can act as a blackboard. The blackboard can be updated using the standard Prolog database update predicates (assert/1, retract/1 and friends).
Waiting is implemented using a POSIX condition variable and matching mutex. On a matching database change the condition variable is signalled using a broadcast, waking up all threads waiting in thread_wait/2. Multiple database updates can be grouped and cause a single wakeup using thread_update/2. This predicate also allows signalling the module condition variable without updating the database and controlling whether all or a single thread is activated.
The blackboard architecture is a good match for an intelligent agent system that has to react on a changing world. Input threads gather sensor data from the world and update a shared world view in a set of dynamic predicates in one or more modules. Agent threads listen to this data or a subset thereof and trigger actions. This is notably a good match with tabling, in particular incremental tabling (see section 7.7) and Well Founded Semantics (see section 7.6).205Future versions may provide additional triggers, for example to learn about invalidated tables. Please share your experience.
The wait is associated with a module. This module is derived from the Options argument.
The Options list specifies when Goal is re-evaluated and optionally when the call terminates due to a timeout.
wait_preds(+Preds)
is not provided.+(PI)+p/1
is read as /(+(p),1)., Goal is only triggered if
a clause was added (assert/1).
If the element is -(PI)wait_preds
option that have been modified.
List must be unbound at entry.The execution of Goal is synchronized between all threads calling this predicate on the same module, changes to dynamic predicates in this module and calls to thread_update/2 on the same module.
This predicate raises a permision_error exception when
called recursively or called from inside a transaction. See
section 4.14.1.2 for
details about interaction with transactions.
broadcast,
default) or a single thread (signal).
Compatibility The thread_wait/2
predicate is modelled after the
Qu-Prolog
predicate thread_wait_on_goal/2.
It is largely compatible. Our current implementation does not support
predicate time stamps.208See predicate_property/2,
property generation. We made this predicate
act on a specific module rather than the entire database. The timeout
specification follows that of the other thread waiting predicates and
may be combined with the retry_every option. The default
retry-time is also 1 second rather than infinite.
The predicates in this section provide signalling between threads. A thread signal inserts any goal as an interrupt into the control flow of any target thread. The target thread processes the goal at the first safe opportunity. The mechanism was introduced with two goals in mind: (1) running a goal inside a thread for debugging purposes such as enabling the status or get access thread-specific data and (2) force a thread to abort its current goal by inserting an exeption into its control flow.
Over time, more complicated use cases have been identified that may result in multiple signals that occur (nearly) simultaneous. As of version 8.5.1 the interface has been extended and the interaction with other built-in predicates has been specified in much more detail.
ThreadId executes Goal as an interrupt at the first opportunity. Defined opportunities are:
sig_atomic. Currently this only applies
to sig_atomic/1.FALSE after PL_handle_signals()
returned -1, indicating an exception was raised.SIGUSR2 to the signalled thread while the handler is an
empty function. This causes most blocking system calls to return with EINTR.
See also the commandline option
--sig-alert. On Windows, PL_handle_signals()
is called when the user processes Windows messages.If one or more signals are queued, the queue is processed. Processing the queue skips signals blocked due to sig_block/1 and stops after the queue does not contain any more non-blocked signals or processing a signal results in an exception. After an exception, other signals remain in the queue and will be processed after unwinding to the matching catch/3. Typically these queued signals will be processed during the Recover goal of the catch/3. Note that sig_atomic/1 may be used to protect the recovery goal.
The thread_signal/2
mechanism is primarily used by the system to insert debugging goals into
the target thread (tspy/1, tbacktrace/1,
etc.) or to interrupt a thread using e.g., thread_signal(Thread,
abort). Predicates from library library(thread) use
signals to stop workers for e.g. concurrent_maplist/2
if some call fails. Applications may use it, typically for similar
purposes such as asynchronously stopping tasks or inspecting the status
of a task. Below we describe the behaviour of thread signalling in more
detail. The following notes apply for
Goal executing in ThreadId
The call port of sig_atomic/1 does not handle signals. This may notably be used to prevent interruption of the catch/3 Recover goal. For example, we may ensure the recovery goal of a timeout is called using the code below. Without this precaution another signal may run before writeln/1 and raise an exception to prevent its execution. Note that catch/3 should generally not be used for cleanup of resources in case of an exception and thus it is typically fine if its Recover goal is interrupted. Use setup_call_cleanup/3 or one of the other predicates from the call_cleanup/2 family for cleanup.
...,
catch(call_with_time_limit(Time, Goal),
time_limit_exceeded,
sig_atomic(writeln('Time limit exceeded'))).
Besides queues (section 10.3.1) threads can share and exchange data using dynamic predicates. The multithreaded version knows about two types of dynamic predicates. By default, a predicate declared dynamic (see dynamic/1) is shared by all threads. Each thread may assert, retract and run the dynamic predicate. Synchronisation inside Prolog guarantees the consistency of the predicate. Updates are logical: visible clauses are not affected by assert/retract after a query started on the predicate. In many cases primitives from section 10.4 should be used to ensure that application invariants on the predicate are maintained.
Besides shared predicates, dynamic predicates can be declared with the thread_local/1 directive. Such predicates share their attributes, but the clause list is different in each thread.
Thread-local dynamic predicates are intended for maintaining thread-specific state or intermediate results of a computation.
It is not recommended to put clauses for a thread-local predicate into a file, as in the example below, because the clause is only visible from the thread that loaded the source file. All other threads start with an empty clause list.
:- thread_local
foo/1.
foo(gnat).
DISCLAIMER Whether or not this declaration is appropriate in the sense of the proper mechanism to reach the goal is still debated. If you have strong feelings in favour or against, please share them in the SWI-Prolog mailing list.