Report on Workshop Advances in High Performance Scientific Computing
Report by Wojciech Smigaj (UCL).
The workshop Advances in High Performance Scientific Computing on 16 January 2012 in Manchester gathered around 50 participants from all around the UK. The four speakers gave their views on diverse aspects of the present move towards new multicore architectures.
Stan Scott (Queens University Belfast) stressed the need for testing and verification of numerical software developed for high-performance platforms, highlighting the importance of round-off error arising in massive calculations and citing the CADNA library as a valuable tool in diagnosing precision loss. In a visionary talk, George Constantinides (Imperial College London) presented the potential of field-programmable gate arrays for the creation of custom architectures tailored for specific numerical algorithms, such as FFT or iterative solvers. He also discussed the gains and challenges brought by alternative (non-floating-point) machine representations of real numbers. Simon McIntosh-Smith (University of Bristol) outlined the current trends in the design of processor architectures, including the emergence of CPUs with integrated GPUs, and presented the industry’s prognostics on the features of future devices, stressing the tendency for making them more and more heterogeneous. He cited the MAGMA library as an intensively developed numerical tool tailored for heterogeneous systems. Finally, Peter Jimack (University of Leeds) reported on the development of a parallel adaptive multilevel finite-difference solver for the highly nonlinear PDEs governing alloy solidification, thus demonstrating that adaptive methods can be successfully ported to parallel architectures.Slides for all the talks are available.