Many modelling applications target systems with a broad range of dynamical timescales. If only a fraction of the modelled system requires small timesteps, large speedups in processing can be achieved by integrating each part of the system with a local timestep. In the astrophysical simulation example taken here with a range of 214 in timesteps, speed-ups over a single global timestep of a factor of 10 have been achieved in parallel with a particle tree-code. In this regime assumptions about dominant costs and ideal load balancing schemes derived from analysis of single stepping simulations break down. In particular, book-keeping and data management tasks can overtake scientific calculation costs. This work examines a new approach based on associating data with similar timesteps rather than using locality in simulation space to control processing costs and to improve load balance and scalability in parallel.