Efficiency is important, because memory is expensive. Sometimes you don't have enough to hold the entirety of all the programs the machine is running, especially if you are using a large program like an X server. To get around this, Unix uses a technique called virtual memory. It doesn't try to hold all the code and data for a process in memory. Instead, it keeps around only a relatively small working set; the rest of the process's state is left in a special swap space area on your hard disk.

Note that in the past, that “Sometimes” last paragraph ago was “Almost always” — the size of memory was typically small relative to the size of running programs, so swapping was frequent. Memory is far less expensive nowadays and even low-end machines have quite a lot of it. On modern single-user machines with 64MB of memory and up, it's possible to run X and a typical mix of jobs without ever swapping after they're initially loaded into core.

Actually, the last section oversimplified things a bit. Yes, programs see most of your memory as one big flat bank of addresses bigger than physical memory, and disk swapping is used to maintain that illusion. But your hardware actually has no fewer than five different kinds of memory in it, and the differences between them can matter a good deal when programs have to be tuned for maximum speed. To really understand what goes on in your machine, you should learn how all of them work.

The five kinds of memory are these: processor registers, internal (or on-chip) cache, external (or off-chip) cache, main memory, and disk. And the reason there are so many kinds is simple: speed costs money. I have listed these kinds of memory in increasing order of access time and decreasing order of cost. Register memory is the fastest and most expensive and can be random-accessed about a billion times a second, while disk is the slowest and cheapest and can do about 100 random accesses a second.

Here's a full list reflecting early-2000 speeds for a typical desktop machine. While speed and capacity will go up and prices will drop, you can expect these ratios to remain fairly constant — and it's those ratios that shape the memory hierarchy.

We can't build everything out of the fastest kinds of memory. It would be way too expensive — and even if it weren't, fast memory is volatile. That is, it loses its marbles when the power goes off. Thus, computers have to have hard disks or other kinds of non-volatile storage that retains data when the power goes off. And there's a huge mismatch between the speed of processors and the speed of disks. The middle three levels of the memory hierarchy (internal cache, external cache, and main memory) basically exist to bridge that gap.

Linux and other Unixes have a feature called virtual memory. What this means is that the operating system behaves as though it has much more main memory than it actually does. Your actual physical main memory behaves like a set of windows or caches on a much larger "virtual" memory space, most of which at any given time is actually stored on disk in a special zone called the swap area. Out of sight of user programs, the OS is moving blocks of data (called "pages") between memory and disk to maintain this illusion. The end result is that your virtual memory is much larger but not too much slower than real memory.

How much slower virtual memory is than physical depends on how well the operating system's swapping algorithms match the way your programs use virtual memory. Fortunately, memory reads and writes that are close together in time also tend to cluster in memory space. This tendency is called locality, or more formally locality of reference — and it's a good thing. If memory references jumped around virtual space at random, you'd typically have to do a disk read and write for each new reference and virtual memory would be as slow as a disk. But because programs do actually exhibit strong locality, your operating system can do relatively few swaps per reference.

It's been found by experience that the most effective method for a broad class of memory-usage patterns is very simple; it's called LRU or the “least recently used” algorithm. The virtual-memory system grabs disk blocks into its working set as it needs them. When it runs out of physical memory for the working set, it dumps the least-recently-used block. All Unixes, and most other virtual-memory operating systems, use minor variations on LRU.

Virtual memory is the first link in the bridge between disk and processor speeds. It's explicitly managed by the OS. But there is still a major gap between the speed of physical main memory and the speed at which a processor can access its register memory. The external and internal caches address this, using a technique similar to virtual memory as I've described it.

Just as the physical main memory behaves like a set of windows or caches on the disk's swap area, the external cache acts as windows on main memory. External cache is faster (250M accesses per sec, rather than 100M) and smaller. The hardware (specifically, your computer's memory controller) does the LRU thing in the external cache on blocks of data fetched from the main memory. For historical reasons, the unit of cache swapping is called a line rather than a page.

But we're not done. The internal cache gives us the final step-up in effective speed by caching portions of the external cache. It is faster and smaller yet — in fact, it lives right on the processor chip.

If you want to make your programs really fast, it's useful to know these details. Your programs get faster when they have stronger locality, because that makes the caching work better. The easiest way to make programs fast is therefore to make them small. If a program isn't slowed down by lots of disk I/O or waits on network events, it will usually run at the speed of the smallest cache that it will fit inside.

If you can't make your whole program small, some effort to tune the speed-critical portions so they have stronger locality can pay off. Details on techniques for doing such tuning are beyond the scope of this tutorial; by the time you need them, you'll be intimate enough with some compiler to figure out many of them yourself.

Even when you have enough physical core to avoid swapping, the part of the operating system called the memory manager still has important work to do. It has to make sure that programs can only alter their own data segments — that is, prevent erroneous or malicious code in one program from garbaging the data in another. To do this, it keeps a table of data and code segments. The table is updated whenever a process either requests more memory or releases memory (the latter usually when it exits).

This table is used to pass commands to a specialized part of the underlying hardware called an MMU or memory management unit. Modern processor chips have MMUs built right onto them. The MMU has the special ability to put fences around areas of memory, so an out-of-bound reference will be refused and cause a special interrupt to be raised.

If you ever see a Unix message that says “Segmentation fault”, “core dumped” or something similar, this is exactly what has happened; an attempt by the running program to access memory (core) outside its segment has raised a fatal interrupt. This indicates a bug in the program code; the core dump it leaves behind is diagnostic information intended to help a programmer track it down.

There is another aspect to protecting processes from each other besides segregating the memory they access. You also want to be able to control their file accesses so a buggy or malicious program can't corrupt critical pieces of the system. This is why Unix has file permissions which we'll discuss later.