ABSTRACT
Today’s trend is for modern hardware architectures to offer increasing numbers of processors and cores. Sutter [2005] wrote, ‘The free lunch is over’ as many of the traditional approaches to improving performance ran out of steam. Energy costs, and the difficulty of dissipating that energy, have led hardware manufacturers away from increasing clock speeds (power consumption is a cubic function of clock frequency) towards placing multiple processor cores on a single chip (where the increase in energy consumption is linear in the number of cores). As there is no reason to expect this trend to change, designing and implementing applications to exploit the parallelism offered by hardware will become more and more important. On the contrary, heterogeneous and other non-uniform memory architectures will only increase the need for programmers to take the particular characteristics of the underlying platform into account.
