First-principles molecular dynamics (FPMD) codes are used by biologists, chemists, and physicists to calculate interactions between atoms and molecules. These codes are based on quantum mechanics modeling and many elements of these computations are proportional to the square or cube of the system size. Thus, with traditional algorithms, the size of simulations is limited to a few hundred atoms, which is much too small for modeling complex systems or realistic materials.
LLNL researchers are developing a new algorithm with reduced complexity and better parallel scaling to enable FPMD simulations with tens of thousands of atoms or more. With this algorithm, the number of atoms that can be simulated is directly proportional to the number of processors available. Using the new algorithm and a new FPMD code called MGmol, they have achieved excellent scaling on 100,000 cores of Vulcan with 100,000 atoms, computing results at a rate of about 4 minutes per time step. Solving the same problem using QBox, an FPMD code implementing a more traditional algorithm, would take several hours.