Student Posters Repository

To address the task scheduling challenges in irregular applications on GPU systems, I propose a new task scheduler that utilizes task parallelism. It incorporates a work stealing strategy to dynamically balance the workload across GPUs. Additionally, I will integrate this task scheduler into the Tascell language, enabling efficient GPU computing with a relatively small programming effort. 

To evaluate the performance, I will apply this proposal to various irregular applications such as hierarchical matrices (H-matrices) construction and arithmetic, sparse matrix applications. Specifically, for H-matrices arithmetic, I will integrate the GPU-supported version into HACApK, an open-source library for H-matrices. The improved implementations of HACApK and Tascell will be made available on GitHub. 

Dense bacterial suspensions show distinctive turbulent motions with an intrinsic length scale. When such dynamics is restricted within some confinement or by some obstacles, bacteria show different behaviors near the boundaries. If these geometrical constraints possess a length scale comparable to that of turbulent motion, they rectify the collective motion of bacteria and yield stably ordered vortices.

We perform agent-based simulations to uncover the underlying mechanism of this ordered behavior and to understand the effect of curved geometrical constraints on the dynamics of bacteria. We model bacteria as self-propelled rods with excluded volume so that they have anisotropic interaction with each other and with boundaries. We investigate the dynamics of bacteria within confinement or with obstacles.

This work uses high-performance computing in a number of ways. Each of the simulation techniques involves the simulation of thousands, to hundreds of thousands of atoms and their interactions. For molecular dynamics simulations, the highly parallelized (MPI + OPENMP/CUDA) LAMMPS and GROMACS molecular dynamics engines were used. For the QM/MM calculations, the CHARMM-GAUSSIAN interface was used, and was likewise parallelized. In addition to the actual simulations, analysis of simulations requires parallel programming as the output of molecular simulations is typically a time-dependent trajectory of the coordinates of each atom in the system. Thus, calculations of observables like the water hydration around specific components of the membrane requires efficient calculation of pariwise distances. Right now, I have acheived this through a combination of Fortran and OpenMP programming; however, it is my longterm goal to shift to a GPU-based approach for these type of analyses. 

Topology optimization is a method used to determine the optimal shape of an object to maximize its desired performance. It is garnering attention as a pivotal technology in the realm of future manufacturing.

Topology optimization has primarily advanced within the field of solid mechanics since its proposal by Bendsøe and Kikuchi in 1988. In recent years, it has been expanded to encompass diverse design problems, including fluid dynamics, heat transfer, acoustics, as well as coupled multi-field issues.

My research focuses on applying topology optimization to address unsteady thermal and fluid problems. In the field of thermal design, topology optimization techniques hold significant appeal as they play a crucial role in controlling the behavior of heat and fluid through appropriate product geometry. This control is essential for enhancing equipment performance, such as the heat exchange and dissipation capabilities of heat transfer devices. Since it is necessary to repeatedly perform unsteady thermal-fluid analysis, the amount of calculation becomes enormous. Therefore, utilization of HPC-related technology is desirable.

Topology optimization is an innovative design method that can maximize the desired mechanical performance based on mathematical theory. In the field of fluid flow topology optimization, many studies assume steady-state conditions. For unsteady flows, it requires a huge computational cost and is difficult to compute with a sufficiently fine mesh. On the other hand, the parallel performance of computers keeps improving in recent years, and it is essential to develop a numerical method with high parallel efficiency that takes full advantage of such performance. Based on these backgrounds, my research aims to develop an unsteady flow topology optimization method suitable for massively parallel computing. In my progress to date, I have developed a numerical program and executed numerical examples of over 10 million cells. In the future, I plan to measure the parallel efficiency (weak scaling) of the proposed method and extend it to multi-physics topology optimization considering coupled fluid-solid interaction phenomena.

Millisecond pulsars (MSPs) are laboratories to study matter in the densest state as well as test general relativity and other relativistic theories of gravity. Their exquisite rotational stability, as measured from arrival time analyses, also allows them to be used as Galactic scale gravitational wave detectors. Finding and studying more MSPs is crucial for optimally using them to study the aforementioned fundamental physics. A very effective way to do this is by generating realistic models of the MSP population. I use Markov-Chain Monte-Carlo techniques to optimize the parameters for MSP population synthesis. These are computationally intensive simulations perfectly suited to be parallelized on HPC. I have also led a search for MSPs using radio observations. These searches were parallelized and conducted on a 12,000 core HPC due to the broad parameter space that needs to be explored to find MSPs. 

Millisecond pulsars attain their short periods by accreting matter from a companion. This is why to fully understand MSP population, binary orbital dynamics must also be taken into account for population synthesis. To date, this has not been done. This is partially due to the computational cost associated with this task. I aim to tackle this problem using HPC

Flexoelectricity is an electromechanical two-way physical coupling between electric polarization and strain gradients (direct flexoelectricity) or between polarization gradients and strain (converse flexoelectricity). It originates at the nanoscale, in non-uniform deformation settings such as bending, torsion or inhomogeneous tension/compression. It is found in all dielectrics, regardless of their microscopic crystalline structure, becoming magnified at small length scales. It has several applications in nanotechnology and engineering, including non-piezoelectric N/MEMS, nanogenerators, ultra-high storage density memories, metamaterials, etc. However, its modeling, quantification, design and exploitation is very challenging, and it is convenient to tackle them via both continuum and quantum mechanics.

The aviation industry is pushing to achieve net zero CO2 emissions by 2050 while sustaining a 4-5% annual increase in passenger transport. The difficulty of this endeavour, along with the limited room for improvement in traditional tube-and-wing aircraft, has stimulated research on unconventional aircraft configurations such as the blended wing-body (BWB). Fortunately, advances in HPC have rendered high-fidelity computational fluid dynamics (CFD) tractable in aerodynamic shape optimization. For example, by distributing different regions of a fine mesh across multiple processors, we can efficiently calculate the drag over the aircraft at each design iteration and alter its geometry to minimize fuel burn. However, drag is a function of weight, motivating the need to explore the tradeoff between the two metrics. To date, researchers have combined CFD and finite element analysis (FEA) to perform aerostructural optimization on conventional aircraft, but studies on unconventional aircraft remain mixed-fidelity due to the computational cost of coupled simulations. Low-fidelity models neglect or oversimplify important physical phenomena, leading to suboptimal geometries. To fill this gap, we apply high-fidelity aerostructural optimization to a BWB with stability and control requirements to generate a more credible estimate of its performance.