This chapter may be skipped on a first reading. Its purpose is to give the advanced user a better understanding of how the system uses memory resources. In a high level language like Prolog it is often not obvious for the programmer to see where the system allocates or frees memory.
The sizes of the different memory areas can be queried by means of the predicate statistics/2 and statistics/0 prints a summary of all these data. Here is a sample output:
[eclipse 1]: statistics. times: [1.12, 0.09, 2.74] seconds session_time: 2.74 seconds event_time: 2.74 seconds global_stack_used: 1936 bytes global_stack_allocated: 4456448 bytes global_stack_peak: 4456448 bytes trail_stack_used: 64 bytes trail_stack_allocated: 262144 bytes trail_stack_peak: 4456448 bytes control_stack_used: 564 bytes control_stack_allocated:262144 bytes control_stack_peak: 262144 bytes local_stack_used: 492 bytes local_stack_allocated: 262144 bytes local_stack_peak: 262144 bytes shared_heap_allocated: 1613824 bytes shared_heap_used: 1411000 bytes private_heap_allocated: 73728 bytes private_heap_used: 36992 bytes gc_number: 1 gc_collected: 23472.0 bytes gc_area: 23560 bytes gc_ratio: 99.6264855687606 % gc_time: 0.0 seconds dictionary_entries: 3252 dict_hash_usage: 2117 / 8192 dict_hash_collisions: 314 / 2117 dict_gc_number: 2 dict_gc_time: 0.01 seconds
The used-figures indicate the actual usage at the moment the statistics built-in was called. The allocated value is the amount of memory that is reserved for this area and actually occupied by the ECLiPSe process. The peak value indicates what was the maximum allocated amount during the session. In the following we will discuss the six memory areas mentioned. The gc-figures are described in section 20.2.
The heap is used to store a variety of data:
- compiled code:
- The heap is used to store compiled Prolog code. Consequently its size is increased by the various compile-predicates, the assert-family and by load/1. Space is freed when single clauses (retract) or whole predicates (abolish) are removed from the system. Note that space reclaiming is usually delayed in these cases (see trimcore/0), since the removed code may still be under execution. Erasing a module also reclaims all the memory occupied by the module’s predicates.
- non-logical storage:
- All facilities for storing information across backtracking use the heap to do so. This includes the handle-based facilities (bags, shelves) as well as the name-based facilities (records, non-logical variables and arrays). As a general rule, when a stored term is overwritten, the space for the old value is reclaimed. All memory related to a non-logical store is reclaimed when the store is destroyed (e.g., using erase_array/1, erase_all/1, bag_abolish/1, shelf_abolish/1).
- The dictionary is the system’s table of atoms and functors. The dictionary grows whenever the system encounters an atom or functor that has not been mentioned so far. The dictionary shrinks on dictionary garbage collections, which are triggered automatically after a certain number of new entries has been made (see set_flag/2). The dictionary is designed to hold several thousand entries, the current number of entries can be queried with statistics/0,2.
- various descriptors:
- The system manages a number of other internal tables (for modules, predicates, streams, operators, etc.) that are also allocated on the heap. This space is reclaimed when the related Prolog objects cease to exist.
- When streams are opened, the system allocates buffers from the heap. They are freed when the stream is closed.
- allocation in C-externals:
- If third party libraries or external predicates written in C/C++ call malloc() or related C library functions, this space is also allocated from the heap. It is the allocating code’s responsibility to free this space if it becomes unused.
Note that the distinction between shared and private heap is only relevant for parallel ECLiPSe systems, where multiple workers share the shared heap, but have their own private heap and stacks.
The Local Stack is very similar to the call/return stack in procedural languages. It holds Prolog variables and return addresses. Space on this stack is allocated during execution of a clause and deallocated before the last subgoal is called (due to tail recursion / last call optimisation). This deallocation can not be done when the clause exits nondeterministically (this can be checked with the debugger or the profiling facility). However, if a deallocation has been delayed due to nondeterminism, it is finally done when a cut is executed or when execution fails beyond the allocation point. Hence the ways to limit growth of the local stack are
The main use of the Control Stack is to store so-called choicepoints. A choicepoint is a description of the system’s state at a certain point in execution. It is created when more than one clause of a predicate apply to a given goal. Should the first clause fail, the system will backtrack to the place where the choice was made, the old state will be restored from the choicepoint and the next clause will be tried. Disjunctions (;/2) also create choicepoints.
The only way to reduce Control Stack usage is to avoid unnecessary nondeterminism. This is done by writing deterministic predicates in such a way that they can be recognised by the system. The debugger can help to identify nondeterministic predicates: When it displays an *EXIT port instead of EXIT then the predicate has left a choicepoint behind. In this case it should be checked whether the nondeterminism was intended. If not, the predicate can often be made deterministic by
The Global Stack holds Prolog structures, lists, strings and long numbers. So the user’s selection of data structures is largely responsible for the growth of this stack (cf. 5.4). In coroutining mode, delayed goals also consume space on the Global Stack. It also stores source variable names for terms which were read in with the flag variable_names being on. When this feature is not needed, it should be turned off so that space on the global stack is saved.
The global stack grows while a program creates data structures. It is popped only on failure. ECLiPSe therefore provides a garbage collector for the Global Stack which is called when a certain amount of new space has been consumed. See section 20.2 for how this process can be controlled. Note again that unnecessary nondeterminism reduces the amount of garbage that can be reclaimed and should therefore be avoided.
The Trail Stack is used to record information that is needed on backtracking. It is therefore closely related to the Control Stack. Ways to reduce Trail Stack consumption are
The Trail Stack is popped on failure and is garbage collected together with the Global Stack.