Improving Locality of Unstructured Mesh Algorithms on GPUs

Abstract text:

Improving Locality of Unstructured Mesh Algorithms on GPUs

András Attila Sulyok1, Gábor Dániel Balogh1, István Zoltán Reguly1 and Gihan R. Mudalige2

1 Pázmány Péter Catholic University, Information Technology and Bionics
2 The University of Warwick, Department of Computer Science


To most efficiently utilize modern parallel architectures, the memory access
patterns of algorithms must make heavy use of the cache architecture:
successively accessed data must be close in memory (spatial locality) and
one piece of data must be reused as many times as possible (temporal

Unstructured mesh algorithms are notoriously difficult in this sense,
especially due to computations that indirectly modify data, leading to race
conditions. In this work we address this problem through a number of
optimisations on GPUs, specifically the use of the shared memory and a
two-layered colouring strategy to cache the data. We also look at different
block layouts to analyse the trade-off between data reuse and the amount of

We developed a standalone library that can transparently reorder the
operations done and data accessed by a kernel, without modifications to the
algorithm by the user. Using this, we performed measurements on relevant
scientific kernels from different applications, such as Airfoil, Volna,
Bookleaf, Lulesh and miniAero; using Nvidia Pascal and Volta GPUs. We
observed significant speedups (1.2--2.5x) compared to the
original codes.

András Attila Sulyok

End of talk: 6/21/2018 12:00:00 PM

Start of talk: 6/21/2018 11:40:00 AM