Section: Scientific Foundations

The context

Technological context: The advent of multi- and many- cores architecture

For almost 30 years after the introduction of the first microprocessor, the processor industry has been driven by the Moore's law till 2002, delivering performance doubling every 18-24 months on a uniprocessor. However since 2002 and despite new progress in integration technology, the efforts to design very aggressive and very complex wide issue superscalar processors have essentially been stopped due to poor performance return, as well as power consumption and temperature walls.

Since 2002-2003 the microprocessor industry has followed a new path for performance: the so-called multicore approach, i.e., integrating several processors on a single chip. This direction has been followed by the whole processor industry. At the same time, most of the computer architecture research community has taken the same path, focusing on issues such as scalability in multicores, power consumption, temperature management and new execution models, e.g. hardware transactional memory.

In terms of integration technology, the current trend will allow to continue to integrate more and more processors on a single die. It will probably be technologically feasible to double the number of cores every two years for the next 12 or 15 years, thus potentially leading to more than a thousand processor cores on a single chip. The computer architecture community has coined these future processor chips as many-cores.

The application context: multicores, but few parallel applications

For the past five years, small scale parallel processor chips (hyperthreading, dual and quad-core) have become mainstream in general-purpose systems. They are also entering the high-end embedded system market. At the same time, very few (scalable) mainstream parallel applications have been developed. Such development of scalable parallel applications is still limited to niche market segments (scientific applications, transaction servers).

The overall picture

Up to now, the end-user of multicores is experiencing improved usage comfort because he/she is able to run several applications at the same time. Eventually, in the near future with the 8-core or the 16-core generation, the end-user will realize that he/she is not experiencing either a functionality improvement or a performance improvement on current applications: the end-user will then realize that he/she needs more effective performance rather than more cores. He/she will then ask either for parallel applications or for more effective performance on sequential applications.