This paper describes a new fast and implicitly parallel approach to
neighbour-finding in multi-resolution Smoothed Particle Hydrodynamics (SPH)
simulations. This new approach is based on hierarchical cell decompositions and
sorted interactions, within a task-based formulation. It is shown to be faster than
traditional tree-based codes, and to scale better than domain decomposition-based
approaches on hybrid shared/distributed-memory parallel architectures, e.g.
clusters of multi-cores, achieving a 40× speedup over the Gadget-2 simulation code.