Optimistic parallel computation: an example from computational chemistry

Abstract

Performance penalties due to synchronization are a common concern in parallel programming. Traditional approaches enforce the correct ordering of write operations using locks, but this can be time-consuming and drastically reduce the benefits of using a parallel machine. Instead, for certain classes or programs we propose using an optimistic approach where the solution is calculated without any locks. This approach detects data races by maintaining statistics on memory writes and correcting potentially inappropriate data values by repeating selected computations and write operations. This scheme is evaluated with a novel parallel implementation of the Moller-Plesset perturbation theory energy calculation for closed-shell molecules.