Seminar on Multicore Programming




Moore's law of doubling every 18 months is still valid in terms of transistor density, but has failed for several recent years in terms of processor speed. In fact, in recent years the greatest increase processing speed has been accomplished by integrating several CPUs on the same chip. This trend will undoubtedly continue and there already are commercial processors with tens or even hundreds of computational cores.

In this seminar we shall investigate current and near-term hardware technologies in multicore systems (e.g. Intel/AMD, Sun) as well as general purpose computing on graphics coprocessors (e.g. NVIDIA), and software technologies (e.g. OpenMP, MPI, Cilk++, CUDA) that allow the programmer to use these new resources. We also review both easily and hardly parallellizable algorithms. Passing the seminar requires frequent attending, a seminar presentation and a small programming exercise on a multicore and/or GPU.

Timo Lilja
Last modified: Thu Mar 5 13:41:47 EET 2009