HIGH PERFORMANCE AGENT-BASED MODELS WITH REAL-TIME IN SITU VISUALIZATION OF INFLAMMATORY AND HEALING RESPONSES IN INJURED VOCAL FOLDS

Abstract

The introduction of clusters of multi-core and many-core processors has played a major role in recent advances in tackling a wide range of new challenging applications and in enabling new frontiers in BigData. However, as the computing power increases, the programming complexity to take optimal advantage of the machine's resources has significantly increased. High-performance computing (HPC) techniques are crucial in realizing the full potential of parallel computing. This research is an interdisciplinary effort focusing on two major directions. The first involves the introduction of HPC techniques to substantially improve the performance of complex biological agent-based models (ABM) simulations, more specifically simulations that are related to the inflammatory and healing responses of vocal folds at the physiological scale in mammals. The second direction involves improvements and extensions of the existing state-of-the-art vocal fold repair models. These improvements and extensions include comprehensive visualization of large data sets generated by the model and a significant increase in user-simulation interactivity.

We developed a highly-interactive remote simulation and visualization framework for vocal fold (VF) agent-based modeling (ABM). The 3D VF ABM was verified through comparisons with empirical vocal fold data. Representative trends of biomarker predictions in surgically injured vocal folds were observed. The physiologically representative human VF ABM consisted of more than 15 million mobile biological cells. The model maintained and generated 1.7 billion signaling and extracellular matrix (ECM) protein data points in each iteration. The VF ABM employed HPC techniques to optimize its performance by concurrently utilizing the power of multi-core CPU and multiple GPUs. The optimization techniques included the minimization of data transfer between the CPU host and the rendering GPU. These transfer minimization techniques also reduced transfers between peer GPUs in multi-GPU setups. The data transfer minimization techniques were executed with a scheduling scheme that aims to achieve load balancing, maximum overlap of computation and communication, and a high degree of interactivity. This scheduling scheme achieved optimal interactivity by hyper-tasking the available GPUs (GHT). In comparison to the original serial implementation on a popular ABM framework, NetLogo, these schemes have shown substantial performance improvements of 400x and 800x for the 2D and 3D model, respectively. Furthermore, the combination of data footprint and data transfer reduction techniques with GHT achieved high-interactivity visualization with an average framerate of 42.8 fps. This performance enabled the users to perform real-time data exploration on large simulated outputs and steer the course of their simulation as needed.