Student Posters Repository

Magnetic fields play a significant role in galaxy evolution because they have similar energy densities to the fluid. Unfortunately, magnetohydrodynamical (MHD) simulations of galaxies are incredibly computationally expensive to run to the point where MHD simulations of entire galaxies have been impossible to perform at high resolution. My current work is to add MHD to Cholla. Cholla is a massively parallel, GPU accelerated, fluid dynamics code that is capable of harnessing the power of GPU accelerated supercomputers such as Summit and Frontier; the Cholla group is also one of the early access teams for Frontier. I also built a testing framework for Cholla that utilizes GoogleTest and is designed with testing on exascale systems in mind. This testing framework supports unit, integration, and system tests and integrates easily with automated testing services such as GitHub Actions and Jenkins.

A classical novae is an astrophysical event, in which a white dwarf accretes the material from its less dense companion star. After enough material is piled up on the white dwarf surroundings, a thermonuclear runaway occurs, ejecting CNO with other nuclides into the interstellar medium. The astronomical observations suggest the accreted material and the outer white dwarf atmosphere layer may mix under specific mechanisms and instabilities. In our present work, we use CASTRO, which stands for Compressible Astrophysics Hydrodynamics, to explore how this mixing process is achieved under the assumption of an initial density, temperature, pressure, and an initial composition profile. Our code is embedded on AMReX that refines our hydrodynamic grid calculations using the Berger & Rigoutsos clustering algorithm, supported by MPI, OpenMP, and GPU parallelization techniques.

The purpose of this poster is to outline efforts at TRIUMF dedicated to testing fundamental symmetries of the Standard Model from nuclear theory. The CKM top--row unitarity test is one such test where nuclear theory can have impact, as the largest contributor, V_ud, can be extracted from super-allowed Fermi beta decays. For precision SM tests, we require theoretical corrections to the observed experimental transitions, and recent analysis of these corrections suggests a discrepancy with unitarity on the order of 2 to 3 sigma. There are two nuclear structure dependent corrections to the Fermi transition values: (i) the isospin symmetry breaking correction and the weak nuclear structure correction. Using the bound state NCSM approach, and the more advanced extension NCSMC which incorporates the continuum, we can evaluate these corrections from ab initio theory.

The Cooley-Tukey Fast Fourier Transform (FFT) algorithm is considered one of the top ten algorithms of the 20th century, and parallel FFT libraries are constantly being developed for applications ranging from astrophysics, molecular dynamics, signal processing,  and machine learning. In this poster we present the most recent developments and scalability results of parallel FFT on GPU accelerators from three vendors: AMD, Intel and NVIDIA. We show visualization results to explore the well-known communication bottleneck, and describe auto-tuning techniques for faster computation and linear-scaling towards exascale machines.  We also present FFT Benchmark performance experiments of state-of-the-art FFT libraries, with an emphasis on those supporting hardware of large-scale heterogeneous systems with accelerators that are available in pre-exascale systems. The results are obtained using an FFT Benchmark harness that was developed to easily benchmark and compare numerous FFT libraries on the US Department of Energy exascale systems.

Digital devices infiltrate many aspects of our daily lives: 39% of Canadians use digital devices to track their health and nearly 60% of American smartphone users use their devices for health-related purposes. By using digital devices, we leave traces that contain information about our cognitions, emotions, and behaviors. Insights gained from smartphones and connected wearable gadgets tell us about social interactions, mobility patterns, and stress reactivity as we go about our daily lives. Personal sensing allows us to quantify trends from these data, holding promise for scaling population health screening, improving early detection and management of disease, and connecting people with evidence-based support. My research focuses on improving health and wellbeing during childhood. What personal sensing features signal risk for cardiovascular disease and psychological distress among parents and children, and how can we use personal sensing to identify targets to intercept the progression of disease? Can we use these insights to connect families with evidence-based interventions that will allow them to gain control of their health and wellbeing? Under this research program, I aim to develop frameworks for i) constructing models using active and passive smartphone and wearable data to identify early-warning signals of precursors to cardiovascular disease and psychological distress in parents and children, ii) situating these patterns within complex systems models to advance theory and simulate intervention feasibility, and iii) linking less invasive, portable measures from digital gadgets to more expensive brain signals to refine the location of functional abnormalities that can then be targeted by medication, psychotherapy, and other clinical therapeutics. In addition to helping make technology smarter so these digital devices can give more accurate information about people’s physical and mental wellbeing, these objectives will inform the design of scalable digital health solutions that increase access to healthcare and evidence-based support in real-world contexts. To achieve these research aims, we first need to develop cloud computing solutions that can securely store passive sensing data that can be leveraged in mobile sensing and wearable technology studies. 

There are currently no reliable mechanisms to determine the power consumption of processes or hardware devices in Linux. This project aims to close the gap.
Key contributions:
1) Understanding why power consumption is a first-class design constraint in HPC systems
2) Demonstrating the value of power consumption across different stakeholders, including end-users, application programmers, kernel developers, and system designers

3) A deeper understanding of power constraints driving modern system design, and a future perspective on application development for energy-efficient HPC platforms

We simulate the motion of a hyperelastic sphere in a background Schwarzschild spacetime using a finite element discretization and a Lagrangian formulation of the equations of motion. We set the initial spacetime coordinates and velocities of the nodes of the discretized sphere on a constant coordinate time hypersurface such that the tidal and elastic forces are balanced. We compute the coordinates and velocities of the center of mass in a Fermi normal frame (FNFnode) centered about a fiducial node. We then compute a geodesic that starts out with the same spacetime coordinates and velocities as the center of mass. Next, we construct a Fermi normal frame (FNFgeo) that is carried along the geodesic. The metric in each Fermi normal frame is nearly flat assuming that the size of the sphere is small compared to the curvature of spacetime. We simulate a close encounter orbit as well as a radial plunge, and observe the sphere as it deforms, oscillates and rotates in the FNFgeo. We observe how the center of mass, spin and elastic energy change as the sphere interacts with the black hole.

 The goal to overcome the prohibitive cost of large eddy simulation and the unreliability of Reynolds-averaged Navier-Stokes predictions under extreme grid features and $Re$ variations, mesh partitioning and distribution, and dynamic mode interplay has generated a new variety of hybrid turbulence modeling methods for turbulent flows.  As a new machine learning technique, the continuous eddy simulation (CES) method integrates a learning strategy for the hybridization of  modeling-focused and grid-resolution-focused methods. It is applied to well-known model structures (e.g. Spalart-Allmaras type equations, which is a typical  two-equation models) as well as other hybridization types and in various computational versions. Unlike traditional hybrid techniques, the new grid-driven resolving approach ensure communication and balancing of resolved and modeled motions and correctly reflect the physics of separated turbulent flows under extreme Reynolds numbers.  For the computational validation of the new applied method, we present the numerical solutions realized through different CES mode variations. And based on the dynamic mode redistribution mechanism, the new simulation methods are found feasible and provide a reliable prediction of very high $Re$ number turbulent flow, which is a challenging problem for existing techniques.  The Computational results are compared with the experimental measurement data of the NASA's 2004 Computational Fluid Dynamics Validation Workshop on Synthetic Jets and Turbulent Separation Control.

Quantum chemistry represents real molecules through atomic nuclei and electrons around them. Electrons distribution is described with algebraic 3D basis functions called atomic orbitals. Interaction of atomic orbitals creates chemical bonds in molecules. From the mathematical point of view, such interactions are the result of integration over atomic orbitals (so-called electron integrals) as a part of solving the Hartree-Fock equation. To investigate the presence and strength of the chemical bonds, accurate yet efficient computational approach is required at the ab initio level of theory. We developed an original “Through-space/Through-bond” method to analyze chemical bonds via “cutting” orbital interactions, i.e. deleting electron integrals that correspond to the particular bond. By comparing molecular properties before and after “cutting” the bond, we can estimate its relative significance in energy.

Because the most computationally expensive part of any quantum chemical calculation is a 4-index two-electron integrals treatment, we use MPI in our programming code for efficiency. In particular, we distribute calculations between multi cores during 4-index integration and transformation of 4-index integrals from one basis set to another.

In application of our method, we show how mutations in nucleotides affect chemical bonds in DNA chain and cause mismatches during DNA replication.