Profiling programs

In this example, we will illustrate the use of:

  • events

  • monitors

by defining a simple profiler that prints the starting and ending time for processing a message sent to an object.

Messages as events

In a pure object-oriented system, all computations start by sending messages to objects. We can thus define an event as the sending of a message to an object. An event can then be specified by the tuple (Object, Message, Sender). This definition can be refined by interpreting the sending of a message and the return of the control to the object that has sent the message as two distinct events. We call these events respectively before and after. Therefore, we end up by representing an event by the tuple (Event, Object, Message, Sender). For instance, if we send the message:

| ?- foo::bar(X).

X = 1
yes

the two corresponding events will be:

(before, foo, bar(X), user)
(after, foo, bar(1), user)

Note that the second event is only generated if the message succeeds. If the message as a goal have multiple solutions, then one after event will be generated for each solution.

Events are automatically generated by the message sending mechanisms for each public message sent using the (::)/2 operator.

Profilers as monitors

A monitor is an object that reacts whenever a spied event occurs. The monitor actions are defined by two event handlers: before/3 for before events and after/3 for after events. These predicates are automatically called by the message sending mechanisms when an event registered for the monitor occurs. These event handlers are declared as public predicates in the monitoring built-in protocol.

In our example, we need a way to get the current time before and after we process a message. We will assume that we have a time object implementing a cpu_time/1 predicate that returns the current CPU time for the Prolog session:

:- object(time).

    :- public(cpu_time/1).
    :- mode(cpu_time(-number), one).
    ...

:- end_object.

Our profiler will be named stop_watch. It must define event handlers for the before and after events that will print the event description (object, message, and sender) and the current time:

:- object(stop_watch,
    % event handler predicates protocol
    implements(monitoring)).

    :- uses(time, [cpu_time/1]).

    before(Object, Message, Sender) :-
        write(Object), write(' <-- '), writeq(Message),
        write(' from '), write(Sender), nl, write('STARTING at '),
        cpu_time(Seconds), write(Seconds), write(' seconds'), nl.

    after(Object, Message, Sender) :-
        write(Object), write(' <-- '), writeq(Message),
        write(' from '), write(Sender), nl, write('ENDING at '),
        cpu_time(Seconds), write(Seconds), write(' seconds'), nl.

:- end_object.

After compiling and loading the stop_watch object (and the objects that we want to profile), we can use the define_events/5 built-in predicate to set up our profiler. For example, to profile all messages that are sent to the object foo, we need to call the goal:

| ?- define_events(_, foo, _, _, stop_watch).

yes

This call will register stop_watch as a monitor to all messages sent to object foo, for both before and after events. Note that we say “as a monitor”, not “the monitor”: we can have any number of monitors over the same events.

From now on, every time we sent a message to foo, the stop_watch monitor will print the starting and ending times for the message execution. For instance:

| ?- foo::bar(X).

foo <-- bar(X) from user
STARTING at 12.87415 seconds
foo <-- bar(1) from user
ENDING at 12.87419 seconds

X = 1
yes

To stop profiling the messages sent to foo we use the abolish_events/5 built-in predicate:

| ?- abolish_events(_, foo, _, _, stop_watch).

yes

This call will abolish all events defined over the object foo assigned to the stop_watch monitor.

Summary

  • An event is defined as the sending of a (public) message to an object.

  • There are two kinds of events: before events, generated before a message is processed, and after events, generated after the message processing completed successfully.

  • Any object can be declared as a monitor to any event. A monitor shall reference the monitoring built-in protocol in the object opening directive.

  • A monitor defines event handlers, the predicates before/3 and after/3, that are automatically called by the runtime engine when a spied event occurs.

  • Three built-in predicates, define_events/5, current_event/5, and abolish_events/5, enables us define, query, and abolish both events and monitors.