CMP Seminar - Calculating system properties on-the-fly in DL_POLY
When: Thursday, October 26, 2023, 2:00 PM - 3:00 PM
Where: G. O Jones Building, Room 516, Mile End
Speaker: Harvey Devereux
Performing calculations on-the-fly can confer performance advantages by eliminating the need to save data or perform input/output operations at runtime. Such algorithms operate by sequentially updating their state as new data arrive. A simple example is the running-average where the mean of a series of data can be computed without storing the entire series. This paradigm, also termed ``online algorithms'' [1], is ubiquitous in the analysis of big data streams where only storing data ephemerally is possible.
In molecular dynamics the system trajectory, mapping atomic configurations through simulation time, is an example of such a data stream. However, analysis after the fact remains a common paradigm. Such an approach in large scale systems, e.g., reaching 1 Billion atoms (length scales of 1 µm) is infeasible. By casting standard analysis of system trajectories as online algorithms, we can eliminate trajectory storage. Saving both computation time and storage space at runtime, as well as additional resources used by post-hoc analysis.
In DL_POLY, we have implemented a general framework for calculating arbitrary correlation functions of simulation properties, on-the-fly. This enables us to calculate, for example, sheer-viscosity and thermal conductivity, via sheer-stress and heatflux correlation functions, as well as computing the velocity autocorrelation (VAF). We achieve this by leveraging the established Multi-tau correlator algorithm [3]. Our implementation kernelises the input data to correlators, meaning new correlations may be supported simply by supplying a kernel to obtain the new observable (scalar, vector or matrix). We realise both significant runtime and storage performance increases even with small or modest systems around N~10^3. E.g., a two order of magnitude reduction in storage costs when calculating the VAF for a system of 500 Argon atoms. We also validate our approach by measuring viscosity and thermal-conductivity at runtime.