An Introduction to Garbage Collection
Part I—The Real Costs of C++ Memory Management
by Richard Gillam
Advisory Software Engineer
IBM Center for Java Technology-Silicon Valley

A couple of months ago, I discussed in the Java Liaison column the debate over whether Java can be considered a "serious programming language" when it relies on an object-file format that must be interpreted at runtime, and when it uses a garbage collector to manage memory. I tried to make the case that these two properties aren’t necessarily the evil things they’re often thought to be in the C++ community.

I spent a lot of that time talking about interpreted byte code and comparatively little talking about garbage collection, promising I’d follow up with a feature article later on that explored the issue in more depth. This is that article.

A year ago, when I went to the C++ World conference, I was struck by the amount, and frequently the vehemence, of the anti-Java rhetoric put out by many of the speakers. I guess this is what happens when you threaten the way someone’s always done things. This is fine, but there also seemed to be a disturbing amount of misinformation floating around. I actually heard one speaker claim that programs that used garbage collection would be three times slower than the same programs without it.

Now, there are numerous reasons why a program in Java might be three times slower than the same thing in C++, and there are a lot of organizations throughout the industry (including my own group and many others in other parts of IBM) that are working madly away fixing them. In fact, I think it’s fair to say that the performance of most Java implementations has improved significantly since I attended C++ World a year ago, and will improve even more in the next year to eighteen months.

But even in the early implementations of Java where it was fair to accuse Java of being three times slower than C++, it was patently ridiculous to blame all, or even most, of this differential on garbage collection. It was only one contributor. To further make the leap that garbage collection, by its very nature, must always impose a 300% execution-time penalty on programs and languages that use it is beyond ridiculous. This simply isn’t true.

For a long time, I’ve seen garbage collection as this mysterious area of computer programming where something magical happens under the covers that prevents the programmer from having to worry about memory leaks and other bugaboos. Some time ago, I decided to do some research on the topic and try to relieve some of the mystery and find out just what happens inside a garbage collector. This turned out to be a fascinating exercise, and I learned a lot and had some of my preconceptions about memory management exploded along the way.

I suspect that there are a lot of programmers out there who know as little about garbage collection as I did when I started my research. What I’d like to do here is present an introduction to garbage collection, presenting the subject in enough detail to give you a good idea of what the various tradeoffs and advantages between various techniques are and a good idea of what goes on under the hood in a typical garbage collector. My purpose here is not to sell either Java or garbage collection, but simply to inject some reality into the debate.

This subject is large enough that it’s really impossible to do a good job of covering it in a single article. This month, we’ll take a long look under the hood of a typical manual memory management subsystem and at how reference counting can be used to help manage memory automatically. Next month, we’ll dive headfirst into the ins and outs of the various tracing methods of automatic memory management and how their various tradeoffs are usually balanced.

The three basic kinds of memory allocation

Let’s start at the beginning: Probably the single most valuable computing resource that an application makes use of is memory. Even in today’s world of multi-megabyte RAM configurations and virtual memory, memory is still a scarce and precious resource whose use must be carefully controlled. When a program is launched, the operating system parcels out to it some amount of memory for its use. The application is then responsible for using this memory to do its job. A program will quickly use up this memory unless it takes care to recycle blocks of memory when the data they contain is no longer needed by the program—that is, a program must take advantage of the fact that very few objects persist from program launch until program termination and re-use the same locations in memory for many different objects whose lifetimes don’t overlap.

In C++, there are basically three methods of allocating memory to store things. The simplest of these is static allocation, which is used for storing global variables and variables declared static (both local static variables and static class members). These variables have to persist for the entire run of a program—there is no point where they "go out of scope." In static storage, a certain range of memory is set aside for the storage of static variables, and this memory can never be used for anything else for the duration of the program’s execution. Each variable name is associated with a memory address somewhere in this space. There is no runtime cost associated with static storage, but static space is permanently allocated and thus not available for anything else.

Then there’s stack allocation, which is used for storing function parameters and non-static local variables. Depending on the processor or OS architecture, either the hardware or the application program maintains a stack, which is used for storing these things. When a function is invoked, it reserves enough space on the stack for its parameters and local variables. This space is reclaimed when the function terminates. Stack space is really cheap in runtime performance: allocating a function’s local variables (its stack frame) is simply a matter of adding or subtracting an appropriate value from the current address of the top of the stack (the stack pointer). Reclaiming a function’s stack frame when the function terminates is equally cheap: you restore the stack pointer to its original value.

The OS or the application has to set aside a fixed amount of space for the stack. Even though the stack can grow or shrink, potentially allowing the space to be used for some other purpose when the stack doesn’t need it, this isn’t done, because then you’d have to be careful to move things out of the way when the stack expanded, which is an unreasonable cost. But if the stack exceeds the amount of memory set aside for it (stack overflow), you’re just stuck and the program terminates. In a virtual-memory system, typically a really huge amount of address space is set aside for the stack, which reduces the likelihood of stack overflow. But if a program has many simultaneous threads of execution, this can cause its own problems. Just the same, since any individual stack frame only persists for as long as a particular invocation of a function is in effect, stack storage is pretty efficient. You typically don’t have to worry about stack overflow except in rare cases of really long call chains, really deep recursion, or a function trying to allocate many unreasonably big objects (say, a bunch of 10,000-character strings) on the stack. (Typically, most programmers see a stack overflow only when they have infinite recursion, which of course is a bug in the first place.)

Finally, there’s heap allocation. The heap is all of the memory allocated to a program that isn’t being used to store the stack, the static variables, or executable code. Typically, this is most of the program’s address space. As a rule, C++ doesn’t automatically allocate anything on the heap. To put something on the heap, you ask the memory manager for a block of memory of a given size. The memory manager will then satisfy this request with memory from the heap. You ask for a block of memory with malloc() in C or with the new operator in C++. In other words, the heap is the pool of memory that’s available to satisfy calls to malloc(). (Actually, the runtime also puts things into the program’s address space, so some proportion of "the heap" actually isn’t available to satisfy calls to malloc()).

You use the heap (the new operator) to store objects that you want to persist longer than a single function call but shorter than the whole program’s duration, and to store larger objects and more complex data structures.

Unlike stack storage, heap storage isn’t automatically reclaimed; you have to do that by hand. If you forget, the heap will eventually fill up, leading to crashes or to severely compromised functionality. You allow storage to be reclaimed (and thus reused to satisfy a subsequent new) by calling free() in C or using the delete operator (which usually calls free()) in C++. If you religiously delete objects when you’re done with them, heap storage can be pretty efficient, but of course, this can be difficult, as it isn’t always obvious when you’re really done with something.

Heap allocation is considerably slower than stack allocation, both because of the extra bookkeeping the memory manager must do, and often because of the extra bookkeeping the programmer must do.

new and delete

It’d be really nice, of course, if you didn’t have to worry about reclaiming memory. You could just do new to your heart’s content and not have to worry about anything. And new could be made quite fast, since it wouldn’t have to worry about reclaiming the memory later.

Of course, you can’t really work this way. Even in today’s world of 64 or 128 Mb RAM configurations and multi-gigabyte disk drives, you’ll eventually start to run low on memory. Computer programs use a lot of memory on a temporary basis—if none of that ever went away, you’d find yourself exhausting even these vast reserves of storage space surprisingly quickly. And you’d be running into serious performance problems long before you actually ran out of memory.

So you have to do delete. Let’s take a look at what actually goes on when you use new and delete. The basic idea is that the memory management subsystem (in this case, the implementations of malloc and free) has to maintain some kind of list of what memory is available and what memory is in use by the application. It also has to have some way of identifying the locations of the various blocks of memory that the application has allocated so it knows what memory has been made available by a delete.

This second job is handled by appending extra bytes to the beginning of every block of memory returned by new. These extra bytes are called a block header, and they carry information about the block. In our example, let’s assume that the only thing about the block that we need to keep track of is its size. On most modern processors, things work faster if they’re stored on long-word boundaries, so let’s say the block header is four bytes long. You’d practically never run into a memory block that actually needs four bytes to store its size, but you want the actual value to be long-word aligned (a three-byte block header would cause real havoc on most 32-bit machines), so we’ll use four bytes.

Now if every block is tagged with a length, you can fairly easily walk across all the allocated blocks. You’d have the base address of the heap stored in static space somewhere; that gives you the address of the first block. Then you can calculate the address of each succeeding block from the address and size of the block before it. This gives you a way to walk all the blocks in the heap.

The next thing we need to deal with is which blocks are free and which are in use. One way of doing this is to maintain a linked list of free blocks (the free list). A pointer to the first one (and thus to the whole free list) would also be stored in static space somewhere. The first four bytes of the block (after the block header) would contain a pointer to the next free block, and so on until the last free block, which would have a null pointer. Remember that the block of memory is not in use by the program, so it’s okay to write the free-list link over whatever data might have been there. (You don’t want to write over the block header because the size of the block is very important.) And if we always allocate things in multiples of four bytes (again, for performance reasons on 32-bit processors), you know you’ve always got enough space in the block for this pointer. The smallest block that could be allocated would be four bytes.

So now you have a pointer to a linked list of free blocks. When you call new, it searches the free list for a block that is large enough to satisfy the request for memory. If a block is too big, it satisfies the request by dividing it into two blocks—it takes away from the block the amount of memory requested by the program, and leaves the remainder on the free list. In the process, four bytes of the block are lost to become the new block’s block header. When you do a delete, the delete operator simply adds the block to the free list, making it available for reuse by a subsequent new.

There are many ways that the free list can be maintained, all of which have different performance tradeoffs. Probably the simplest way to maintain the free list is in order of last deletion. When you delete an object, it simply gets tacked onto the front of the free list. This makes deletion quite fast, but can really slow down allocation. When you allocate a block, you simply walk the free list until you find a block whose size is greater than or equal to the block size the application wants. You have no clues as to how far you’ll have to walk to find such a block. If the first block you find is too big, you go ahead and divide it up, even if there’s a smaller appropriate block one link further down the chain. And, of course, you might walk the whole free chain without finding a block of appropriate size, in which case you’re stuck.

One problem here is that you can divide big blocks to satisfy memory requests, but you don’t have a good way to stitch two adjacent free blocks back together into one big free block. One way around this is to borrow a bit from the block header to indicate whether a block is free or in use. As we showed before, the length is all you need to find the address of the next block in the list, so if you can go straight to that block and ask it if it’s free without having to look for it in the free list, the delete operator can coalesce a free block with the block that follows it in memory. The delete operator cannot, however, coalesce a free block with the block that precedes it in memory, because you have no way to move backward through the heap. However, the new operator can use the "free" bit to detect adjacent free blocks and coalesce them.

It should be fairly obvious that all this work takes time. When you request a block of memory from the heap, the new operator has to walk the free list one block at a time until it finds a block that’s big enough. If it detects adjacent free blocks along the way, it probably also must stop and coalesce them (which involves adjusting the header of the one lower in memory and walking the free list again to delete the one higher in memory from the free list), even if the combined block still isn’t big enough to satisfy the request. You can defer this work until you know you have to coalesce two blocks together, but then you have to walk all the way to the end of the free list and then start over again looking for adjacent free blocks, which could be very time-consuming.

The delete operator here has less work to do, but it still has work to do. Its basic job is to append the block being deleted to the end of the free list, but it should probably also check to see if the block that follows it in memory is also free and coalesce them if possible. Since deleting the block that’s higher in memory from the free list requires another walk through the free list, this can be expensive too.

You can make coalescing blocks together a lot easier by maintaining the free list in address order. This probably doesn’t do much to new’s performance (in fact, it may speed it up by making sure that everything that can be coalesced already is), but it takes more time at delete time. Now delete, rather than being able to slap the deleted block onto the end of the free list, has to walk the free list to find an appropriate place. But if the newly freed block is adjacent to a freed block on either side, combining them into a single block is a snap, and doesn’t require walking the free list again.

You can also minimize the dividing and coalescing that has to happen by maintaining the free list in size order. new still has to walk the whole free list, but it’s guaranteed that the first appropriate free block it encounters will be the smallest free block that will satisfy the request.

You can also balance some of these requirements against each other by using multiple free lists. You could, for example, maintain pools of objects of different sizes. An object would thus move from one free list to another when it gets divided in two or combined with another object. For any given allocation or deletion, however, you wouldn’t have to waste time examining every potential node in the free list. You could also maintain two free lists of all the free blocks in two different orders: say, address order (to make coalescing fast) and size order (to minimize dividing of blocks). Another possibility is to defer deleting of nodes that have been absorbed through coalescing from the free list until later by maintaining a separate list of nodes that have to be deleted in this way. Yet another possibility is to include bidirectional pointers so that you could traverse the free list in either direction. The problem with all of these approaches is that they require more space for the free-list links, effectively raising the minimum block size and wasting more memory. They also may cost more in extra bookkeeping time than they save.

You could also use an array instead of a linked list for the free list. Then you could binary-search the free list for appropriate blocks. The big problem here is that you couldn’t use the free blocks themselves to store the free list, meaning you’d have to set aside an extra area of memory just for the free list. And you’d have to deal somehow with the possibility that that extra area of memory might not be big enough.

In reality, most commercial memory managers are much more complicated than the simple examples we’ve examined here, but they all must deal with the same basic sets of tradeoffs.

Problems of manual memory management

Obviously, however you organize the free list(s), new and delete are potentially expensive operations, and therefore C++ programmers are trained to minimize their calls to new and delete as much as possible. This tends to mean preferring stack-allocated objects over heap-allocated objects whenever possible (which is a good idea unless you do it so much you cause a stack overflow).

It also means C++ programmers often resort to tricks to minimize creations and deletions. One such trick is to maintain a separate pool of objects of some frequently-used class. Once you’ve created an object of that class, it stays around. Instead of deleting it, you add it to an internally-maintained free list and then re-use it the next time you need an object of that class (a variation on this keeps just one extra object of that class around, which is useful if you go through lots of complete new/delete cycles in succession with the same class). In effect, when you do this kind of thing, you’re saying the C++ memory manager won’t do your job well enough, and so you’re going to write your own memory manager that is more suitable for your purposes. You’re probably right, but you also wind up doing a lot of extra work And you leave yourself open to errors in your memory-management code, which probably hasn’t been tested as thoroughly as the C++ memory manager.

In addition to the performance problem, there are other problems with normal C++ memory management. When the free space is interspersed with blocks of memory that are in use, you have fragmentation. Fragmentation can lead to situations where there’s enough free memory to satisfy a request for memory, but no single free block is big enough, causing a program to run out of memory before it strictly has to (and degrading the performance of the memory manager before that). This kind of memory-management system can also lead to poor locality of reference. Locality of reference has to do with the relative positions in memory of objects that refer to each other. If objects that point to each other are close to each other in memory, that’s good locality of reference, and if objects that point to each other are widely separated in memory, that’s bad locality of reference. In a flat memory model, locality of reference doesn’t matter. But with a virtual-memory operating system, or a cached processor architecture, bad locality of reference leads to more page faults and cache misses, with a potentially huge impact on performance.

In addition to all of these problems, manual memory management is prone to programmer error. You can get into trouble by deleting an object too soon (dangling pointers, which can lead to all manner of ugly problems), deleting an object twice (double deletion, which can lead to cycles in the free list, with obvious consequences), or not deleting an object at all (memory leaks, which can degrade your program’s performance and lead to crashes and/or badly compromised functionality when memory gets low). Avoiding these problems in a program of any complexity can be quite ugly, and is easy to get wrong. C++ programmers spend an inordinate amount of time debugging memory-related problems (which are often notoriously hard to track down), and practically every commercially-available software product written in C++ has memory leaks. Furthermore, C++ programs often have to do quite a bit of bookkeeping to avoid these memory bugs, meaning the runtime and memory-usage cost of manual memory management is actually usually bigger than just that incurred by the memory manager itself.

Sure, memory management is hard. But does it have to be this hard? This is where garbage collection comes in.

Reference counting

Obviously, many people have had this thought. It’d be really nice if we could just new up objects as we needed to and leave it to "the system," somehow, to know when to release them. This would save the program from extra bookkeeping hassles and from the errors that can happen when you get memory management wrong. In other words, it’d be really nice if we had some form of automatic memory management.

The simplest and most obvious form of automatic memory management is a technique known as reference counting. In fact, many, if not most, C++ programmers have used reference counting at one time or another when it wasn’t feasible to do manual memory management any other way. This is very frequently the "extra bookkeeping" programmers have to do in order to get memory management right. So if instead of the programmer doing reference counting, the system did it, the programmer would be freed of this burden and also wouldn’t have to worry about deleting objects.

The basic idea behind reference counting is this: each object contains a count of how many other objects point to it. This is called (surprise, surprise) a reference count. Whenever a pointer is set to point to the object, the reference count goes up by one, and whenever a pointer that points to the object gets set to point to something else, or goes away altogether, the reference count goes down by one. When the object’s reference count reaches zero, you know nothing in the program can access the object, and therefore the program’s done with the object and it’s safe to delete it.

Very simple and easy to implement. Works great for a lot of applications. But this technique has a number of problems. The first problem is the extra overhead for maintaining the reference count. We’ve done nothing to reduce the cost of newing and deleting objects, and to it we’ve added the additional cost of maintaining the reference count. Now every time you assign to a pointer, instead of just simply copying the pointer value, you also have to adjust two objects’ reference counts, and possibly delete one or more objects. If either object is paged out of memory, you have to bring it in from disk, which will increase the cost more (of course, you’d always have to page the object in from disk when you actually access the object, but you generally don’t do that nearly as often as you manipulate pointers to it). This is a pretty serious extra cost. The nice thing about manual memory management is that although you have to implement reference counting on your own, you only have to pay the price for it when you absolutely need it.

Then we have the problem of where to store the reference count. The obvious answer is in the block header. In the simplest implementation, this would balloon the block header to eight bytes. Twice the overhead required by simple manual memory management.

There are ways of reducing this overhead. We can keep the size of the block header at four bytes by reducing the number of bytes we use for the block size. Two bytes would cover most rationally-sized blocks, leaving us with two bytes for the reference count. But that limits us to 64 Kb blocks, which might be kind of small. Instead, we might opt for a three-byte size, leaving us with a byte for the reference count.

A one-byte reference count imposes a maximum of 255 references to an object, which ought to be plenty. But what happens if you exceed this limit? In that case, you’re stuck. You have nowhere to store the count of additional references, so you can never reduce the reference count as references to the object go away. This object now can never be deleted.

Of course, you can get around this by combining reference counting with one of the more sophisticated memory-management techniques we’ll talk about next month. In fact, this works pretty well.

There’s even a refinement that can be made: Most objects only have a reference count of one. So you can cover most of your cases with a one-bit reference count. The bit would mean either "one" or "many." When a reference to an object whose reference count was "one" went away, the object could be immediately deleted. Once the reference count went to "many," the object would have to be handled with tracing garbage collection. The beauty of this is that it reduces the number of objects that have to be handled by generalized garbage collection, and it’s easy to find space for the one-bit "reference count." In fact, in one-bit implementations, the bit is usually in the pointers themselves, not in the block header. It also greatly reduces the amount of overhead that is required to maintain the reference count. Because the "reference count" is stored in the individual pointers rather than the block header, you don’t have to page in the actual object if it’s paged out, and adjusting the reference count when you’re adding a reference to an object involves simply setting a single bit. (You don’t even have to look at that bit first, because it doesn’t really matter whether the object was already shared or not.)

Lumpy deletion

There’s another limitation of reference counting, and this problem is generally shared by manual memory management as well. One complaint you hear often about garbage collection is that the garbage collector can run at unpredictable times and that the whole program grinds to a halt while the garbage collector is running. This not only wastes time, but it does so in a particularly obtrusive way. With both manual memory management and reference counting, the argument goes, you have complete control over when memory management happens, it generally takes less time, and what time it does take is amortized over the whole run of the program, not done in separate passes that halt all the action.

This is the common wisdom about garbage collection, and like most "common wisdom," it contains a certain amount of truth, but it’s not nearly as black and white as all that. Modern garbage collectors can do a pretty good job of spreading their cost more evenly across a program’s execution time, and reference counting and manual memory management don’t spread things out anywhere near as evenly as most people think. Programmers tend to think they’re in complete control of when memory management happens and that its cost is spread out evenly across the program. Neither of these things is generally true (although both could be if you were willing to put in the extra effort). Instead, people are comfortable with manual memory management because it involves a lot less mystery. You’re not ceding a lot of control to this mysterious black box that you didn’t write and trusting it to handle your memory problems for you; instead, all the work is right there where you can see it. Unfortunately, one generally pays a heavy price for this lack of mystery, if not in runtime efficiency, then definitely in ease of design, coding, and maintenance.

In normal C++ memory management, every object has an owner, which is the object or function that is responsible for deleting it. That object’s destructor includes a delete for every object that it owns. The cost of the delete (which is not trivial, as we saw above, because of the challenge of maintaining the free list in a particular order and of coalescing free blocks together) would be smoothly amortized across the run of the program if news and deletes were randomly distributed. But they rarely are. Instead, you tend to have clusters of related objects. A single object may have several objects that it owns, which means all those objects go down at the same time as the owning object does. If they, in turn, own more objects, the problem snowballs. Collections are particularly bad this way: typically there’s one owning reference to the whole collection and when that goes away, so do all the objects in the collection. So a single delete (or a single smart pointer going out of scope) could result in a massive number of deletes and destructor calls, if the collection is large. Distribution of news and deletes is rarely smooth; it’s "lumpy." And most programmers don’t do anything to smooth out the "lumpiness"; they just get rid of things as soon as they know they’re done with them. So unless you’re taking extra action to smooth out these lumps, you really are exerting very little control over when you pay the price of memory management. It happens when it happens, and the cost is not trivial.

This problem affects both manual memory management and most reference-counting schemes. It’s possible, even reasonable, to design a reference-counting scheme that eases this problem, but it’s more complicated.


Finally, there’s the big Achilles’ heel of reference counting: cycles. Typical reference counting schemes can’t handle cycles in the reference chain— that is, self-referential data structures. These can come up disturbingly often. If you have circularly linked lists, bidirectional linked lists, or back pointers (for example, items in a collection that refer back to the collection itself), you have cycles. If you have a vector of foo, and foo contains a pointer back to the vector that contains it, then a vector of ten foos has a reference count of 10 before any outside references are taken into account. Even when the last outside reference to the vector goes away, meaning the vector and all the foos can go away, the vector still has a reference count of 10 (and all the foos still have reference counts of at least 1).

In a typical C++ program whose author has implemented reference counting by hand, this kind of thing can be avoided: the programmer, who understands the semantics of the objects in the program, can bless only certain pointers as ones that affect the reference count (for example, by letting them be instances of some smart-pointer class); the others are all aliases and can go away without a problem (for example, they’re just raw C++ pointers). On the other hand, with a generalized reference-counting mechanism built into a language’s runtime, this generally isn’t possible—the memory manager doesn’t know anything about the semantics of the pointers or objects it’s manipulating, so it has no way to distinguish between pointers that should affect the reference count and ones that should not.

The language could, of course, be designed to require the programmer to tell it which pointers are which—for example, through the use of a special keyword—but this complicates the language syntax and leaves languages that otherwise could manage memory completely automatically open to the same kinds of memory bugs that languages like C++ with manual memory management are susceptible to. In practice, this is rarely, if ever, done.

Another approach is to combine reference counting with some other form of automatic memory management. Reference counting thus gets rid of what it can, and a garbage collector runs occasionally to clean up the cyclical structures. This gets us back around to the one-bit scheme discussed earlier, and requires that we delve into the bowels of tracing garbage collection.

Stay tuned…

And this is where I’m going to leave things for now. I’ve probably already babbled on well past my word-count target. We’ve now covered the preliminaries, the information which it’s important to know when looking at tracing garbage collection. What I’ve tried to do here is take an in-depth look at the actual costs of regular C++ memory management, and at reference counting, the simplest way to ease the programmer’s bookkeeping burden and make things more automatic. As you can see, these techniques are far from free; we’re just conditioned to accept these costs. This is important to keep in mind as we look at more sophisticated memory management schemes: Their costs should be compared to the actual costs of doing things the way we’ve always done them, not to zero.

Next month, we’ll jump into tracing garbage collection with both feet and see how it actually works and what it actually costs. That’s where things really start to get interesting. Hope to see you back here then.