Computational chemistry covers many fields such as quantum chemistry, molecular dynamics and so on. In addition to theories and experiments, with the technological development of supercomputer and the increasing of computing power, these fields have become a significant part in chemistry research. Semi-empirical method, ab initio method and DFT method are commonly used in quantum chemistry to study the structure, spectroscopic properties and chemical reactions of chemical molecules. While molecular dynamics, aimed at studying the dynamic properties and reaction pathways of chemical systems, includes subdivisions of first principles molecular dynamics, classic molecular dynamics, mesoscopic molecular dynamics and QM/MM method. The requirements for computing resources vary significantly according to the types of computing systems and computing methods. Generally, these software programs have high requirements for CPU computing power, memory capacity and bandwidth, network communication bandwidth and latency, as well as IO read/write bandwidth.
There are numerous software programs in computational chemistry. The characteristics of applications are determined by the type of their underlying algorithms and run scale. Based on different algorithms, the analysis can be categorized as following: ●For software programs that use highly precise ab initio method such as Gaussian, Gamess and MOLPRO, often require a larger memory to store large scale base data sets, and a larger IO bandwidth to read and write temporary files in compute procedures. Hence, it’s helpful to use large memory nodes to meet the memory requirements of these processes, and to choose storage systems with high IO bandwidth to read/write temporary files. ●For NW-Chem, ADF, DFTB+ and many other software programs that use DFT algorithm, reasonable memory sizes are needed to meet the memory bandwidth requirement, because the bottleneck restricting the application acceleration often lies in the capacity and bandwidth of the memory. ●For most molecular dynamics software, especially the classic molecular dynamics including Gromacs and NAMD, a high bandwidth and low latency network is required to accommodate increasing the number of the compute nodes. Alternatively, using acceleration cards to boost computing is also an ideal option as underlying algorithms are relatively simple and most of them already support the acceleration of GPU and MIC cards. In addition, as most molecular dynamics applications are floating-point intensive, using vectorized instruction sets can effectively improve the run efficiency of those applications. The schematic diagram on the right shows the result of three common molecular dynamics software programs optimized with AVX instruction sets.