1.8 Non-Von Neumann Models

Until recently, almost all general-purpose computers followed the von Neumann design. However, the von Neumann bottleneck continues to baffle engineers looking for ways to build fast systems that are inexpensive and compatible with the vast body of commercially available software. Engineers who are not constrained by the need to maintain compatibility with von Neumann systems are free to use many different models of computing. A number of different subfields fall into the non-von Neumann category, including neural networks (using ideas from models of the brain as a computing paradigm), genetic algorithms (exploiting ideas from biology and DNA evolution), quantum computation (previously discussed), and parallel computers. Of these, parallel computing is currently the most popular.

Today, parallel processing solves some of our biggest problems in much the same way as settlers of the Old West solved their biggest problems using parallel oxen. If they were using an ox to move a tree and the ox was not big enough or strong enough, they certainly didn't try to grow a bigger ox-they used two oxen. If a computer isn't fast enough or powerful enough, instead of trying to develop a faster, more powerful computer, why not simply use multiple computers? This is precisely what parallel computing does. The first parallel-processing systems were built in the late 1960s and had only two processors. The 1970s saw the introduction of supercomputers with as many as 32 processors, and the 1980s brought the first systems with over 1,000 processors. Finally, in 1999, IBM announced the construction of a supercomputer called the Blue Gene. This massively parallel computer contains over 1 million processors, each with its own dedicated memory. Its first task is to analyze the behavior of protein molecules.

Even parallel computing has its limits, however. As the number of processors increases, so does the overhead of managing how tasks are distributed to those processors. Some parallel-processing systems require extra processors just to manage the rest of the processors and the resources assigned to them. No matter how many processors are placed in a system, or how many resources are assigned to them, somehow, somewhere, a bottleneck is bound to develop. The best that we can do to remedy this is to make sure that the slowest parts of the system are the ones that are used the least. This is the idea behind Amdahl's Law. This law states that the performance enhancement possible with a given improvement is limited by the amount that the improved feature is used. The underlying premise is that every algorithm has a sequential part that ultimately limits the speedup that can be achieved by multiprocessor implementation.