Speaker
Mr
James Willis
(ICC, Durham University)
Description
The next-generation cosmological code SWIFT has been demonstrated to outperform current codes by more than an order of magnitude on ordinary x86-based clusters such as the cosma system in DiRAC. This has been achieved thank to the use of better algorithm, task-based parallelism and asynchronous communications. However, the next generation of machines that will contain many more cores per node, clocked at lower frequencies, brings more challenges. To recover the same performance than on a regular Xeon-based system, a code has to scale twice as well on a KNL architecture. This implies that all parts of the code have to be well-suited to this paradigm and have to be written in parallel.
In this talk we present some improvements to the SWIFT code aimed at tackling these challenges. Parallelisation of the task scheduler and of most of the operations performed in-between time-steps is now necessary and brings significant improvements to the code scalability on KNL and as a knock-on effect on regular Xeon chips.
Primary author
Mr
James Willis
(ICC, Durham University)
