Logtalk provides experimental support for multi-threading programming on selected Prolog compilers. Logtalk makes use of the low-level Prolog built-in predicates that interface with POSIX threads (or a suitable emulation), providing a high-level set of predicates that allows programmers to easily take advantage of modern multi-processor and multi-core computers without caring about the details of creating, synchronizing, or communicating with threads. Logtalk multi-threading programming integrates with object-oriented programming by allowing the supporting per object threads, enabling objects to call local predicates concurrently and to send and receive asynchronous messages.
Multi-threading support may be disabled by default. It can be enabled on the Prolog configuration files of supported compilers by setting the read-only compiler flag threads to on.
In order to automatically create and set up an object's thread the threaded/0 object directive must be used:
:- threaded.
The thread is created when the object is loaded or created at runtime. The thread for the pseudo-object user is automatically created when Logtalk is loaded (provided that multi-threading programming is supported and enabled for the chosen Prolog compiler).
The message queue associated with an object's thread receives both goals to be proved asynchronously and the replies to the threaded calls made from within the object itself.
Logtalk provides a small set of built-in predicates for running goals in new threads and for retrieving goal results. These predicates provide high-level support for multi-threading programming, covering the most common use cases.
A goal may be proved asynchronously using a new thread by calling the Logtalk built-in predicate threaded_call/1. The term representing the goal is copied, not shared with the thread. Therefore, any variable bindings resulting from the goal proof must be retrieved by calls to the built-in predicate threaded_exit/1.
For example, assume that we want to find all the prime numbers in a given interval, [N, M]. We can split the interval in two parts and then span two threads to compute the primes numbers in each sub-interval:
prime_numbers(N, M, Primes) :-
M > N,
N1 is N + (M - N) // 2,
N2 is N1 + 1,
threaded_call(prime_numbers(N, N1, [], Acc)),
threaded_call(prime_numbers(N2, M, Acc, Primes)),
threaded_exit(prime_numbers(N, N1, [], Acc)),
threaded_exit(prime_numbers(N2, M, Acc, Primes)).
prime_numbers(N, M, Acc, Primes) :-
...
Note that a call to the threaded_call/1 predicate is always true (assuming a callable argument). The threaded_exit/1 calls block execution until the results of the threaded_call/1 calls are sent to the object's thread. In a computer with two or more processors (or with a processor with two or more cores) the code above can be expected to provide better computation times when compared with single-threaded code for sufficiently large intervals.
The results of proving a goal asynchronously in a new thread may be retrieved by calling the Logtalk built-in predicate threaded_exit/1 within the same object where the call to the threaded_call/1 predicate was made.
The threaded_exit/1 predicate allow us to retrieve alternative solutions through backtracking (if you want to commit to the first solution, you may use the threaded_once/1 predicate instead of the threaded_call/1 predicate). For example, assuming a lists object implementing the usual member/2 predicate, we could write:
| ?- threaded_call(lists::member(X, [1,2,3])). X = _G189 yes | ?- threaded_exit(lists::member(X, [1,2,3])). X = 1 ; X = 2 ; X = 3 ; no
Note that, in this case, the threaded_call/1 and the threaded_exit/1 calls are made within the pseudo-object user. The implicit thread running the lists::member/2 goal suspends itself after providing a solution, waiting for the request of an alternative solution; the thread is automatically terminated when the runtime engine detects that further backtracking to the threaded_exit/1 call is no longer possible.
Calls to the threaded_exit/1 predicate block the caller until the object's thread receives the reply to the asynchronous call. The predicate threaded_peek/1 may be used to check if a reply is already available without removing it from the thread queue. The threaded_peek/1 predicate call succeeds or fails immediately without blocking the caller. However, keep in mind that repeated use of this predicate is equivalent to polling a thread queue, which may severely hurt performance. A third predicate, threaded_discard/1, allows us to discard all matching replies available on the object's thread waiting to be retrieved.
Sometimes we want to prove a goal in a new thread without caring about the results. This may be accomplished by using the built-in predicate threaded_ignore/1. For example, assume that we are developing a multi-agent application where an agent may send an "happy birthday" message to another agent. We could write:
threaded_ignore(agent::happy_birthday), ...
The call succeeds with no reply of the goal success, failure, or even exception ever being sent to the object making the call. Note that this predicate implicitly implies a deterministic call of its argument.
The built-in predicate threaded_race/1 allows a disjunction of goals to be interpreted as a set of competing goals, each one running on its own thread. The first thread to terminate successfully leads to the termination of the other threads. This is useful when we have several methods to compute something, without knowing a priori which method will be faster. For example, assume that we have defined a predicate try_method/2 that calls a chosen method, returning its result. We may then write:
method(Result) :-
...,
threaded_race(
( try_method(method1, Result)
; try_method(method2, Result)
; ...
)),
threaded_exit(try_method(Method, Result)),
...
The threaded_exit/1 call will return both the identifier of the fastest method and its result. Note that, when setting up competing goals such as in the example above, we must ensure that the argument of the threaded_exit/1 call unifies with each individual goal in the disjunction used as argument on the threaded_race/1 call. This ensures that the threaded_exit/1 call will return the results of the first method to complete, no matter which one.
Proving a goal asynchronously using a new thread may lead to problems when the goal implies side-effects such as input/output operations or modifications to an object's database. For example, if a new thread is started with the same goal before the first one finished its job, we may end up with mixed output or a corrupted database. In order to solve this problem, predicates (and grammar rule non-terminals) with side-effects can be declared as synchronized by using the synchronized/1 object and category directive. Proving a query to a synchronized predicate (or synchronized non-terminal) is protected by a mutex, thus allowing for easy thread synchronization. For example:
:- synchronized(db_update/1). % ensure thread synchronization
db_update(Update) :- % predicate with side-effects
...
A second example. Assume an object defining two predicates for writing, respectively, even and odd numbers in a given interval to the standard output. Given a large interval, a goal such as:
| ?- threaded_call(obj::odd_numbers(1,1000)), threaded_call(obj::even_numbers(1,1000)). 1 3 2 4 6 8 5 7 10 ... ...
will most likely result in a mixed up output. By declaring the odd_numbers/2 and even_numbers/2 predicates synchronized:
:- synchronized([
odd_numbers/2,
even_numbers/2]).
the second goal will only start after the first one finished:
| ?- threaded_call(obj::odd_numbers(1,1000)), threaded_call(obj::even_numbers(1,1000)). 1 3 5 7 9 11 ... ... 2 4 6 8 10 12 ... ...
Note that, in a more realistic scenario, the two threaded_call/1 calls would be made concurrently from different objects. Using the same synchronized directive for a set of predicates imply that they all use the same mutex, as required for this example.
The synchronized/1 directive must precede any local calls to the synchronized predicate (or synchronized non-terminal) in order to ensure proper compilation. In addition, as each Logtalk entity is independently compiled, this directive must be included in every object or category that contains a definition for the described predicate, even if the predicate declaration is inherited from another entity, in order to ensure proper compilation.
Logtalk supports both deterministic and non-deterministic synchronized predicates (and synchronized non-terminals). However, whenever possible, synchronized predicates should be coded as deterministic predicates in order to avoid deadlocks. In those cases where the predicate (or grammar rule) is defined in the same object (or category) where the predicate is declared synchronized, Logtalk takes advantage of any existing mode/2 directives in order to generate the most appropriated mutex handling code. When no mode/2 predicate directives are presented, Logtalk assumes a deterministic predicate when generating the mutex handling code.
We may declare all predicates of an object (or a category) as synchronized by using the entity directive synchronized/0. In this case, the synchronized/1 predicate directive is not necessary and should not be used.