6.7 Garbage Compactor Pattern
The Garbage Compactor Pattern is a variant of the Garbage Collection Pattern that also removes memory fragmentation. It accomplishes this goal by maintaining two memory segments in the heap. During garbage collection, live objects are moved from one segment to the next, so in the target segment, the objects are juxtapositioned adjacent to each other. The free memory in the segment then starts out as a contiguous block.
The Garbage Collection Pattern solves the problem of programmers forgetting to release memory by every so often finding inaccessible objects and removing them. The pattern has a couple of problems, including maintaining the timeliness of the application and fragmentation. Fragmentation means that the free memory is broken up into noncontiguous blocks. If the application is allowed to allocate blocks in whatever size they may be needed, most applications that dynamically allocate and release blocks will eventually get into the situation where although there is enough total memory to meet the allocation request, there isn't a single contiguous block large enough. At this point, the application fails. Garbage collection per se does not solve this problem just because it finds and removes dead objects. To compact memory, the allocated blocks must be moved around periodically to leave the free memory as a single, large contiguous block.
6.7.3 Pattern Structure
Figure 6-13 shows the structural pattern for copying garbage collection. A copying garbage collector works in a single phase. It is initiated in the same way as a mark and sweep garbage collector. As it searches the object space from the root objects, it copies all the live objects it finds to another memory space. It is more efficient than mark and sweep because a single phase is all that is necessary, and it also eliminates memory fragmentation because it compacts memory as it moves the referenced objects. The copying garbage collector must update object references as it moves objects. This pattern requires twice as much memory as the mark and sweep pattern because it must always maintain both a from space and a to space (although they will reverse roles on subsequent invocations of the garbage collector). In addition, a copying garbage collector requires that user objects reference heap objects via double bufferingthat is, their pointers must point to pointers owned by the heap. This allows the garbage collector to update its internal pointer references to the actual location of the heap object as it moves around. Either that, or the garbage collector must have references to the user objects and modify their pointers in vivo when the referenced heap object is relocated.
Figure 6-13: Garbage Compactor Pattern
6.7.4 Collaboration Roles
The Buffered Ptr is an intermediary between one object's reference to the object being referenced. This is required because the Garbage Compactor must update the references to the objects as it moves them. This is far simpler if the actual references to the memory location are under its control rather than the object's clients.
The Client is the user-defined object that allocates memory (generally, although not necessarily, in the form of objects). During the collection process, root objects are searched, and objects found during the search are moved as they are found.
Free Block List
A list of free blocks from which requests for dynamic memory are fulfilled.
The Garbage Compactor manages the reclamation of memory by searching the object space starting with the root objects and copying the found objects from the current memory segment to the target memory segment. Dead objects are not copied and are thus automatically reclaimed.
The Heap is the owner of all the Segments and Buffered Ptrs. The heap fills all memory requests from the currently active segment and ignores the inactive segment. The roles the two segments play swap each time the Garbage Compactor performs the garbage compaction process.
The Memory Block is just like it sounds: a block (normally of arbitrary size, in which case it contains a size parameter) of memory usually, although not necessarily, associated with an object. Memory Blocks may be pointed to by the Free Block Listin which case they are not currently being pointed to by a Clientor may be pointed to by a Client (via a Buffered Ptr)in which case they are not pointed to by the Free Block List.
The most noticeable consequence of using the Garbage Compactor Pattern is that the programmers don't need to deallocate their objects (the Garbage Compactor does it for them) and that fragmentation does not monotonically increase the longer the system runs. Fragmentation increases for a while but is reduced to zero when the Garbage Compactor runs. Since the Garbage Compactor runs when a request for memory cannot be satisfied, this means that if there is enough total memory to meet a request, the request will always be satisfied.
Another highly noticeable consequence of this pattern, at least in terms of memory requirements, is that the pattern requires twice as much memory as the Garbage Collection Pattern. Assuming that each Segment is large enough to hold the worst-case memory needs at any moment in time, the pattern requires two such segments. This makes this approach inappropriate when there are tight memory size requirements.
Doing pointer arithmetic with the Garbage Compactor Pattern is a chancy thing for a number of reasons. However, since the main reason for doing pointer arithmetic is to manage memory, this should not cause many difficulties.
Of course, the length of time necessary to run the compactor is an issue for any application in which timeliness is a concern. There is a small amount of overhead for the double buffering of the pointers, but the major timeliness impact comes from the time and cycles necessary to identify the live objects and copy them to the target memory segment. This pattern requires more CPU cycles to run than the Garbage Collection Pattern because of the overhead of copying objects, but with care, it may be possible to run the garbage collector piecemeal to implement an incremental garbage compactor at the cost of recomputing which objects in the From segment are still live.
Because the reclaimed objects are not destroyed per se, their destructors are not called. If there are finalizing behaviors required of the objects (other than the normal release of memory), then the programmer must still manually ensure these behaviors are invoked before removing their references to the objects.
6.7.6 Implementation Strategies
There are a number of small details to be managed by the memory management system in this pattern. The use of Buffered Ptrs allows the Garbage Compactor to move the objects and then update the location in a single place. If the source language is interpreted, such as Java, then the virtual machine can easily manage the double pointer referencing required of the client objects (in other words, their pointers are to Buffered Ptrs, which ultimately point to the actual memory used). If the source language produces native code, then the new operator must be rewritten to not only allocate the Memory Block per se but also create a Buffered Pts as well.
6.7.7 Related Patterns
As with the Garbage Collection Pattern, this pattern can seriously impact the predictability of timeliness of systems using it. When timeliness is a primary concern, the Static Allocation, Fixed Sized Buffer, or Smart Pointer Patterns may be better. The Garbage Collection Pattern has the benefit of removing memory leaks, and it requires less memory than the Garbage Compactor Pattern, but it doesn't address memory fragmentation. The Static Allocation and Fixed Sized Allocation Patterns remove or reduce fragmentation but are not immune to memory leaks.
6.7.8 Sample Model
The example shown in Figure 6-14 is the same as for the previous pattern. Figure 6-15 shows a scenario of the system as it collects the garbage. We see that when the Garbage Collector is started, it first requests Segment 2 (the target segment) to initialize itself so that it is ready to begin copying memory into itself. Because it always starts empty, there is no fragmentation within the segment as memory blocks are allocated one after another in a contiguous fashion.
Figure 6-14: Garbage Compactor Pattern
Figure 6-15: Garbage Compactor Pattern Example Scenario
Then the garbage collector searches the root objects. As it finds a root object, it asks if it has any valid links. First, in the case of Ob1, it finds two valid links, so the Garbage Collector can pass the object Ob1 off to the AddObject operation of the target segment (Segment 2). The segment, in turn, gets the location of the buffer pointer for the memory, allocates a new memory block to store Ob1's data, and then copies the object from the original segment. Finally it updates Obj1's buffered pointer to point to its data's new location.