Efficient Use of Computer Resources

Many computational chemistry techniques are extremely computer-intensive. Depending on the type of calculation desired, it could take anywhere from seconds to weeks to do a single calculation. There are many calculations, such as ah initio analysis of biomolecules, that cannot be done on the largest computers in existence. Likewise, calculations can take very large amounts of computer memory and hard disk space. In order to complete work in a reasonable amount of time, it is necessary to understand what factors contribute to the computer resource requirements. Ideally, the user should be able to predict in advance how much computing power will be needed. 

---

In Fig. 4 we compare the timings for three different models, the simple one K per processor, the wrapped algorithm, and a model where two states are assigned per processor sequentially. Note that until J = 50 the one K per processor model job uses the smallest amount of wall clock time. It is clear, however, that this method does not make efficient use of computer resources. The wrapped model, however, scales very well and outperforms the sequential two K per processor model at every / > 0, a clear illustration of the degradation of performance due to load imbalance. 

All the trajectories are reactive. This enhances the efficient use of computational resources. This is in contrast to initial value MD, in which many trajectories do not end at the desired state. 

During the course of the research process, new methods will be applied to analyze and collect information in support of this process. Software systems used in this research must be flexible enough to incorporate new methods into the data and processing models built into the system. The system must support efficient repurposing of compute hardware and flexible description and discovery of computational and data resources within the system. 

Discontinuous molecular dynamics (DMD) simulations can be used to investigate large systems efficiently with moderate computational resources. DMD simulations were designed to be applicable to systems that interact via discontinuous potentials (square-well/square-shoulder and hard-sphere). They proceed by analytically calculating the next collision time. Several papers [26-28] describe the details of DMD simulations. The algorithm complexity of DMD simulations is O (Mog N). (One paper by Paul [29] even claims a realization of the DMD method... 

---