Dissertation paves the way for tomorrow’s supercomputers

Many of today’s server computers contain several processors (the ‘brain’ of the computer) and support for so-called shared memory. Shared memory makes these multiprocessors simple for programmers to use, but it entails great complexity in the actual design of the computer. To achieve speed, a computer’s processor and memory are complemented by smaller ‘cache’ memories that contain the most important computer material. It’s something like having your own little black book to complement the phone book.

With shared memory and several processes, however, the problem arises that every processor’s cache memory must be designed to ‘talk with’ the others so that they will have the same information when a number in the phone book is changed, for instance. Supercomputers with shared memory are built today largely with specially constructed hardware. The great disadvantages of this are long construction times and high prices. Another way to construct supercomputers is to build the shared memory with a layer of software. This entails simpler and cheaper hardware, but also relatively low capacity and poor binary compatibility; in other words, it has not been possible to run all programs with this solution. Therefore, this approach has been considered unworkable since the 1990s.

Håkan Zeffer is now introducing something halfway between these two approaches and a solution to a more than 20-year-old problem in the computer industry. The idea is that the need for ‘coherence’ across memories is detected in the hardware, whereas the actual coordination is executed in the software. The hardware can therefore be not only simple but also small, which makes it possible to use it in many different types of processors. This system is both cheaper and more flexible. Since parts of the system are based on software, the programs and the computer itself can be optimized.

“This is a new way to build inexpensive supercomputers. The complexity is merely a fraction of that of the tradition solution, and the capacity is not only comparable with tradition systems but often better,” says Håkan Zeffer.

The idea has been patented, and the industry has expressed an interest in the new approach.