Multicore computing

As dual-core processors become mainstream in PCs, and the first quad-core desktop systems emerge, there’s little doubt that we are all heading towards a multicore future.

But the academic world is already expressing doubts about how far current processor trends can go before they run into architectural and software problems.

A group of researchers at the University of California-Berkeley’s department of electrical engineering and computer sciences has spent nearly two years considering the implications of multicore and future ‘manycore’ processors. In a recent report, the researchers highlighted the problems – and the opportunities – that lie ahead.

Their report was well timed: in February this year, Intel showed an 80-core processor chip developed as part of its terascale computing research project, capable of a trillion floating-point operations per second.

Although this chip will never go into full production, it’s intended to test the basic concepts of manycore processing and help to explore the problems of software development for massively parallel computing.

Core growth
This might seem of little relevance to typical PC users. Putting a dual-core, or even quad-core, processor in a PC motherboard produces extra throughput without any software changes, as the operating system takes care of scheduling specific tasks for the separate cores and automatically kicks off new threads as processing resources in either core become available.

or the average user, more cores just mean better performance. But as the number of cores rises to eight or 16, the multicore benefits seen in the conventional programming model begin to fall away.

There might be hundreds of threads active whenever you run Windows and seve ral applications concurrently, but most of them are not doing much and there is actually little to be gained by running more than a couple in parallel. A game engine, for example, will need substantial rewriting to take advantage of an eight-core or 16-core processor and a separate graphics co-processor chip.

Getting the best out of future multicore processors won’t be as simple as it might look.

The Berkeley report starts from the premise that multicore chip development has been driven not by technological research but by practical and economic limits on the efficiency of traditional single-processor designs.

As chip geometries shrink and their complexity and clock speeds rise, power dissipation becomes a major concern, and other esoteric problems (such as soft errors caused by cosmic rays, or the increasing effects of signal delay caused by wiring) make development and debugging a slow and difficult process.

This is why there have been diminishing returns in processor performance over the past five years, following a long period when performance figures doubled every 18 months. With uniprocessor development beginning to stall, performance would double only every five years at the present rate.