Why 64-bits are good, and why they're not

THIS ARTICLE hopes to cast some light on why 64-bit addressing, that is, the native mode of the Opteron or Itanium versus that of the Athlon or Pentium is important in 2003. It also attempts to address what the requirements are and - equally importantly - are not.

Before we start, an easy one. Why 64-bit and not 48-bit? Because it costs little more to extend a 32-bit ISA to 64-bit than to only 48-bit, and most people like powers of two. In practice, many of the hardware and operating system interfaces will be less than 64 bits, sometimes as few as 40 bits, but the application interfaces (i.e. the ones the programmers and users will see) will all be 64-bit.

There are several non-reasons quoted on the Internet; one is as arithmetic performance. 64-bit addressing does not change floating-point, and is associated with 64-bit integer arithmetic; while it is easy to implement 32-bit addressing with 64-bit arithmetic or vice versa, current designs don't. Obviously 64-bit makes arithmetic on large integers faster, but who cares? Well, the answer turns out to be anyone who uses RSA-style public key cryptography, such as SSH/SSL, and almost nobody else.

On closer inspection, such use is dominated by one operation (NxN->2N multiplication), and that is embedded in a very small number of places, usually specialist library functions. While moving from 32 to 64 bits does speed this up, it doesn't help any more than adding a special instruction to SSE2 would. Or any less, for that matter. So faster arithmetic is a nice bonus, but not a reason for the change.

File pointers are integers, so you can access only 4GB files with 32 bits, right? Wrong. File pointers are structures on many systems, and files of more than than 4GB have been supported for years on a good many 32-bit systems. Operations on file pointers are usually well localised and are normally just addition, subtraction and comparison anyway. Yes, going to 64-bits makes handling large files in some programs a bit easier, but it isn't critical.

Let's consider the most common argument against 64-bit: compatibility.

Almost all RISC/Unix systems support old 32-bit applications on 64-bit systems, as did IBM on MVS/ESA, and there is a lot of experience on how to do it almost painlessly for users and even programmers.

Microsoft has a slightly harder time because of its less clean interfaces, but it is a solved problem and has been for several decades.

Now let's get onto some better arguments for 64-bit. One is that more than 4GB of physical memory is needed to support many active, large processes and memory map many, large files - without paging the system to death. This is true, but it is not a good argument for 64-bit addressing. The technique that Intel and Microsoft call PAE (Physical Address Extension) allows 64 GB of physical memory but each process can address only 4GB. For most sites in 2003, 64GB is enough to be getting on with.

IBM used this technique in MVS, and it worked very well indeed for transaction servers, interactive workloads, databases, file servers and so on. Most memory mapped file interfaces have the concept of a window on the file that is mapped into the process's address space - PAE can reduce the cost of a window remapping from that of a disk transfer (milliseconds) to that of a simple system call (microseconds). So this is a rather weak reason for going to 64-bit addressing, though it is a good one for getting away from simple 32-bit.

Now, let's consider the second most common argument against 64-bit: efficiency. Doubling the space needed for pointers increases the cache size and bandwidth requirements, but misses the point that efficiency is nowadays limited far more by latency than bandwidth, and the latency is the same. Yes, there was a time when the extra space needed for 64-bit addresses was a major matter, but that time is past, except perhaps for embedded systems.

So 64-bit addressing is unnecessary but harmless except on supercomputers? Well, not quite. There are some good reasons, but they are not the ones usually quoted on the Internet or in marketing blurb.

The first requirement is for supporting shared memory applications (using, say, OpenMP or POSIX threads) on medium or large shared memory systems. For example, a Web or database server might run 256 threads on 16 CPUs and 32GB. This wouldn't be a problem if each thread had its own memory, but the whole point of the shared memory programming model is that every thread can access all of the program's global data. So each thread needs to be able to access, say, 16GB - which means that 32-bit is just not enough.

A more subtle point concerns memory layout. An application that needs 3GB of workspace might need it on the stack, on the main heap (data segment), in a shared memory segment or in memory set aside for I/O buffers. The problem is that the location of those various areas is often fixed when the program is loaded, so the user will have to specify them carefully in 32-bit systems to ensure that there is enough free space in the right segment for when the program needs its 3GB.

Unfortunately, this choice of where to put the data is often made by the compiler or library, and it is not always easy to find out what they do. Also, consider the problem of an administrator tuning a system for multiple programs with different characteristics. Perhaps worst is the case of a large application that works in phases, each of which may want 2GB in a different place, though it never needs more than 3 GB at any one time. 64-bit eliminates this problem.

To put the above into a commercial perspective, almost all general purpose computer vendors make most of their profit (as distinct from turnover) by selling servers and not workstations. 64-bit addressing has been critical for some users of large servers for several years now, and has been beneficial to most of them. In 2003, 64-bit is needed by some users of medium sized servers and useful to most; by 2005, that statement could be changed to say `small' instead of 'medium sized'. That is why all of the mainframe and RISC/Unix vendors moved to 64-bit addressing some time ago, and that is why Intel and AMD are following.

On the other hand, if you are interested primarily in ordinary, single user workstations, what does 64-bit addressing give you today? The answer is precious little. The needs of workstations have nothing to do with the matter, and the move to 64-bit is being driven by server requirements. µ

Nick Maclaren has extensive experience of computing platforms