Latency tolerance through parallelization of time in scientific applications

Emerging computing environments, such as the Grid, promise enormous raw computational power. However, effective use of such platforms is often difficult, because conventional spatial decomposition leads to fine granularity, resulting in high communication overhead. We introduce the concept of guided simulations to parallelize along the time domain. Here, we use the fact that typically results of other simulations of closely related problems are available. In this approach, we automatically and dynamically determine a relationship between old simulations and the one being performed, and use this to parallelize along the time domain. We demonstrate the validity of this approach by applying the technique to an important application involving molecular dynamics simulation of nanomaterials. In this application, spatial decomposition is not effective due to the small size of the physical system. However, time parallelization is effective, since the granularity is much coarser. We also mention how this approach can be extended to make it inherently fault tolerant.