NAMD and molecular dynamics simulations

Molecular dynamics (MD) simulations compute atomic trajectories by solving equations of motion numerically using empirical force fields, such as the CHARMM force field, that approximate the actual atomic force in biopolymer systems. Detailed information about MD simulations can be found in several books such as [1,7]. In order to conduct MD simulations, various computer programs have been developed including X-PLOR [5] and CHARMM [4]. These programs were originally developed for serial machines. Simulation of large molecules, however, require enormous computing power. One way to achieve such simulations is to utilize parallel computers. In recent years, distributed memory parallel computers have been offering cost-effective computational power. NAMD was designed to run efficiently on such parallel machines for simulating large molecules. NAMD is particularly well suited to the increasingly popular Beowulf-class PC clusters, which are quite similar to the workstation clusters for which is was originally designed. Future versions of NAMD will also make efficient use of clusters of multi-processor workstations or PCs.

NAMD has several important features:

• Force Field Compatibility
  The force field used by NAMD is the same as that used by the programs CHARMM [4] and X-PLOR [5]. This force field includes local interaction terms consisting of bonded interactions between 2, 3, and 4 atoms and pairwise interactions including electrostatic and van der Waals forces. This commonality allows simulations to migrate between these three programs.

• Efficient Full Electrostatics Algorithms
  NAMD incorporates the Distributed Parallel Multipole Tree Algorithm (DPMTA) [3] and Distributed Particle Mesh Ewald (DPME) algorithms, which takes the full electrostatic interactions into account. These algorithm reduces the computational complexity of electrostatic force evaluation from $O(N^2)$ to $O(N)$ or $O(N \log N)$.

• Multiple Time Stepping
  The velocity Verlet integration method [1] is used to advance the positions and velocities of the atoms in time. To further reduce the cost of the evaluation of long-range electrostatic forces, a multiple time step scheme is employed. The local interactions (bonded, van der Waals and electrostatic interactions within a specified distance) are calculated at each time step. The longer range interactions (electrostatic interactions beyond the specified distance) are only computed less often. This amortizes the cost of computing the electrostatic forces over several timesteps. A smooth splitting function is used to separate a quickly varying short-range portion of the electrostatic interaction from a more slowly varying long-range component. It is also possible to employ an intermediate timestep for the short-range non-bonded interactions, performing only bonded interactions every timestep.

• Input and Output Compatibility
  The input and output file formats used by NAMD are identical to those used by CHARMM and X-PLOR. Input formats include coordinate files in PDB format [2], structure files in X-PLOR PSF format, and energy parameter files in either  CHARMM or X-PLOR formats. Output formats include PDB coordinate files and binary DCD trajectory files. These similarities assure that the molecular dynamics trajectories from NAMD can be read by CHARMM or X-PLOR and that the user can exploit the many analysis algorithms of the latter packages.

• Dynamics Simulation Options
  MD simulations may be carried out using several options, including
  • Constant energy dynamics,
  • Constant temperature dynamics via
    • Velocity rescaling,
    • Velocity reassignment,
    • Langevin dynamics,
  • Periodic boundary conditions,
  • Constant pressure dynamics via
    • Berendsen pressure coupling,
    • Nosé-Hoover Langevin piston,
  • Energy minimization,
  • Fixed atoms,
  • Rigid waters,
  • Rigid bonds to hydrogen,
  • Harmonic restraints,
  • Spherical or cylindrical boundary restraints.

• Easy to Modify and Extend
  Another primary design objective for NAMD is extensibility and maintainability. In order to achieve this, it is designed in an object-oriented style with C++. Since molecular dynamics is a new field, new algorithms and techniques are continually being developed. NAMD's modular design allows one to integrate and test new algorithms easily. The structure and design of the program is described in the NAMD Programmer's Guide with sufficient detail to allow additions of such new algorithms.

• Interactive MD simulations
  A system undergoing simulation in NAMD may be viewed and altered with VMD; for instance, forces can be applied to a set of atoms to alter or rearrange part of the molecular structure. For more information on VMD, see http://www.ks.uiuc.edu/Research/vmd/.

• Load Balancing
  An important factor in parallel applications is the equal distribution of computational load among the processors. In parallel molecular simulation, a spatial decomposition that evenly distributes the computational load causes the region of space mapped to each processor to become very irregular, hard to compute and difficult to generalize to the evaluation of many different types of forces. NAMD addresses this problem by using a simple uniform spatial decomposition where the entire model is split into uniform cubes of space called patches. An initial load balancer assigns patches to processors such that the computational load is balanced as much as possible. During the simulation, an incremental load balancer monitors the load and performs necessary adjustments.