Parallel Libraries and Domain Specific Frameworks

Parallel algorithms developed for conventional parallel machines are not necessarily appropriate or efficient for petaflops machines. Parallel algorithms for such class of petaflops machines must handle low bisection bandwidth and relatively low memory to processor ratio. They must exploit the availability of dedicated communication threads and the existence of multiple parallel communication links. They must also meet the challenges of scalability.

We have developed and evaluated several parallel frameworks and their applications for petaflops class machines including Molecular Dynamics (MD) and FEM framework.

The molecular dynamics simulation of biomolecules is one of the planned applications for Blue Gene/L. It is a challenging application to parallelize. A microsecond simulation includes about a billion timesteps. Thus, each timestep involves a relatively small amount of computation that must be effectively parallelized.

We have developed an experimental prototype program called LeanMD that models the essential computations of MD. Performance study and analysis of LeanMD on BigSim for Blue Gene/L in [6] shows that LeanMD can scale up to petaflops machines.

In this section, we focus on the parallel FEM framework and its simulation on massive parallel machines.