AutoPas: Auto-Tuning for Particle Simulations

AutoPas in ls1 mardyn: Massively parallel particle simulations with node-level auto-tuning

“…The Array of Structure (AoS) data layout always stores an array (or vector), wherein multiple particles are stored as a structure (or class). The Strucure of Arrays (SoA) data layout will convert the particle data into multiple arrays, one for each property, to allow for better vectorization (refer to [10] for details).…”

“…In this section, we describe the modifications made in ls1 mardyn to support AutoPas and its RegularUpdateInterface in both single-node and multi-node runs. We therefore first shortly discuss ls1 mardyn itself and its previous changes [10]. Then we describe the changes that we implemented in ls1 mardyn to support the new interface of AutoPas in single and multi-node applications.…”

“…As described in Ref. [10], ls1 mardyn was extended with an additional particle class and a wrapper around AutoPas, which replaces the previously used particle container. The particle class hereby implements both the interface needed by ls1 mardyn and the interface needed by AutoPas.…”

“…In preceding work of our group [10], the idea of providing multiple algorithms and then tuning over these algorithms at runtime has been picked up for n-body simulations. We introduced the node-level library AutoPas, integrated it into the highly-parallel MD program ls1 mardyn [11] and demonstrated significant speed-ups for node-level simulations simply by switching between different traversal strategies.…”

“…In the present paper, we are going beyond node-level scope [10] of autotuned MD simulations, considering MD systems distributed on multiple compute nodes. Implications of this extension in contrast to a node-level system are outlined and discussed.…”