An experience report on (auto-)tuning of mesh-based PDE solvers on shared memory systems.

With the advent of manycore systems, shared memory parallelisation has gained importance in high performance computing. Once a code is decomposed into tasks or parallel regions, it becomes crucial to identify reasonable grain sizes, i.e. minimum problem sizes per task that make the algorithm expose a high concurrency at low overhead. Many papers do not detail what reasonable task sizes are, and consider their findings craftsmanship not worth discussion. We have implemented an autotuning algorithm, a machine learning approach, for a project developing a hyperbolic equation system solver. Autotuning here is important as the grid and task workload are multifaceted and change frequently during runtime. In this paper, we summarise our lessons learned. We infer tweaks and idioms for general autotuning algorithms and we clarify that such a approach does not free users completely from grain size awareness