Speaker
Dr
Graeme Stewart
(University of Glasgow)
Description
Modern CPU architectures for scientific computing are characterised by
having multiple CPU cores and wide vector registers. Effective use of
these features requires different design patterns from those that ran
on the single core systems of the past. Instead of serial processing,
parallel processing needs to be at the heart of the new model, with
many operations proceeding concurrently. In this talk I discuss what
that challenge means for scientific code and which programming
patterns are most likely to be useful to scientists. I will look at
some common libraries used to safely implement concurrency and give
some examples from various fields. I will also emphasise the role of
memory layout in achieving good throughput in high performance
scientific code.
