With the availability of petascale computing platforms looming on the horizon, there exists the possibility of simulating complex phenomena involving widely disparate length and time scales on several thousand processors. Such simulations, which involve coupling Eulerian solvers to Particle solvers, are being developed and pursued by researchers in computational fluid dynamics, molecular dynamics and computational astrophysics, to name a few. Since each of the constituent solvers, in such coupled simulations, have their own unique system requirements, the performance of these codes is often quite poor on runs involving a few hundred processors, at present. On petascale machines, the performance of these codes on several tens of thousands of processors is more than likely to be even worse. Scientific codes will, therefore, require considerable re-engineering and algorithmic improvements to take advantage of the emerging petascale platforms that are characterized by deep memory hierarchies and multicore processor
architectures.

This talk will present some of the research problems that the presenter has been involved with and attempts to develop efficient codes with a top-down approach that aims at mapping the algorithm efficiently onto the platform architecture.