Computational steering for spike-coupled neuronal network simulations on high-performance computing resources
Date
2018
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
One significant challenge in neuroscience is understanding the cooperative behavior of large numbers of neurons. Neuroscientists have long postulated that information processing is a result of the synchronicity of spike trains of large groups of connected neurons. How this synchronous behavior produces complex behavior such as planning, language, and emotion is not well understood. Models of brains neuronal networks allow scientists to explore the impact of differential neuronal connectivity using analysis techniques and information not available experimentally. However, encoding more and more biophysically realistic neurobiological models into computational simulations combine increasing computing and data requirements. High-performance computing (HPC) has enabled the simulation of the brain at increasing levels of fidelity; the amount of data produced by simulations on HPC systems is becoming increasingly difficult to save and transform into scientific insights because of the memory/storage characteristics of the systems. ☐ The ability to process, analyze, and produce substantive scientific inferences on data is becoming increasingly difficult. Furthermore, the growth in complexity of brain models makes finding optimal input variable configurations hard, as performing efficient parameter sweeps become impractical due to the increasing number of input parameters and the need of narrowing down the parameters values. Coupling simulation with analysis has been widely investigated in applications such as computational fluid dynamics and other numerical simulations codes, but little work has been done in integrating real-time analysis and interactive steering with brains neural network simulations. Analysis of data produced by brain models is still relegated to traditional post-processing and sequential scientific workflows. ☐ In this thesis, we claim that new workflows that leverage real-time analysis and computational steering methods are required to overcome the increasing divide between computational and I/O subsystem performance. We claim that there exists a need to transform the end-to-end scientific HPC workflow from one that is non-transparent, trial-and-error based, and static to one that is investigative, hypothesis-driven, and adaptive. To this end, we design, implement and evaluate a distributed computational steering environment (CSE) prototype for the GEneral NEural SImulation System (GENESIS) on HPC systems. First, we investigate how the diversification of HPC hardware along with increasing model complexity impact performance and data generation. We show that increasingly biophysically realistic brain models are computationally feasible but require HPC resources to mitigate the burgeoning computing and data requests of the associated simulations. Second, we investigate the integration of an in situ analysis approach with a simulated multi-compartmental model of neocortex. We assess our approach using both qualitative and quantitative methods. We show that by eliminating the need for data movement, using only a local view of our data, results in comparable scientific insights to those that are obtained via a global post-simulation analysis. Third, we design, implement and demonstrate a working prototype of a CSE that leverages data collected at run-time via in situ analysis, to steer a GENESIS simulation in real time. We assess and quantify simulation perturbation (i.e., overhead) and presentation latency of our prototype when steering a simulated neuronal networks models on GENESIS. ☐ The primary goal of this thesis is to show how the use of an environment such as our CSE can support dynamic workflows in brains neural network simulation via computational steering and can mitigate resource usage associated with more naive trial-and-error based approaches. Our work is the first step towards a change in thinking from traditional static, post-simulation to dynamical, adaptive steering simulations of realistic neurobiological models.
Description
Keywords
Biological sciences, Applied sciences, Big data, Computational steering, Data analysis, High performance computing, In situ analysis, Stream analysis