Student Posters Repository

The solution of a large sparse linear system of equations is the most computationally expensive part of most of scientific applications. Our research focuses on boosting performance of sparse linear solvers on distributed memory architectures. In order to improve performances, we have to take into account both the algorithmic and the software engineering aspects. From the algorithmic point of view, we focus on multigrid preconditioners, which have been proved to be particularly efficient. More specifically, we have implemented and tested a matching aggregation algorithms in MLD2P4, a library implementing a great variety of preconditioners. From the software engineering point of view, we are focusing on the migration of the softwares PSBLAS (implementing iterative linear solvers) and MLD2P4 from MPI to Fortran Coarrays (CAF). In order to have the migration happen without bugs-injection, we developed a framework for unit testing for CAF, extending the already existing framework Pfunit. Other than comparing MPI and CAF performances on different communication patterns we aim at describing the best usage strategies when migrating to CAF and at analysing how software quality is affected.

RELION-2 enables autopicking, classification, and refinement on desktops using GPUs. We show that a ~6000$ workstation equipped with four GPUs easily outperforms even a large cluster. Comparing a single CPU cluster node to a GPU workstation of roughly comparable price, speedup ranges from an order of magnitude, for 2D-classification, up to a factor 40 for, 3D-classification(6 classes). For auto-picking the new implementation achieves a  100-fold speedup compared to the previous release. We also show that the new implementations yield qualitatively identical results. This is accomplished by exposing lower levels of parallelism within the algorithm to the massive parallel hardware of the GPUs.

Cardiovascular diseases are amongst the deadliest diseases worldwide according to the World Health Organization (2015), with ischaemia and strokes being the two first causes. This project aims at obtaining accurate predictive tools for cardiovascular diseases, in particular ischaemia. The associated simulations consist of complex electro-mechanical biventricular finite elements (FE) systems of the human heart, in which different drugs and ischaemic conditions are tested.

Heart models comprise of several coupled physical phenomena, which means that electrical propagation, fluid flow and solid deformation interact with each other. The heart geometry is extracted from medical images provided by clinical collaborators (magnetic resonance images, MRI), through a semi-automated process. Due to certain requirements of the electrical propagation, these geometries are discretised into FE systems of tens of millions of nodes. This leads to the need of using HPC platforms to obtain a solution.

Different sets of simulations are used in this study. A first set is used to validate subsequent simulations with the respective 4-dimensional MRI data. Further sets are used in which different ischaemic locations and/or severities are tested, with and without the effect of drugs.

Combining several fields that all require HPC, this will be an overview of 3 separate projects.

Binary Black Hole (BBH) post processing and parameter estimation: BBH are the most likely observable sources of gravitational waves, which are an integral result of general relativity. The LIGO Scientific Collaboration confirmed the existence of gravitational waves and hence BBH in September of 2015 and have continued detection runs. LIGO requires estimates of the waveforms of gravitational waves in order to successfully complete data analysis. These waveforms require numerical relativity and a catalogue of these numerical relativity waveforms is supplied to LIGO through the Simulating eXtreme Spacetimes (SXS) collaboration. SXS requires supercomputers to run BBH simulations and continuously updates their code to reflect advancements in HPC. I run simulations as a member of the SXS, and use the waveforms already in our LIGO catalogue to 1. Analyze and correct for centre of mass drift in the BBH, which affects the reliability of the waveform, 2. Investigate the relationships between inclination frequency, orbital frequency, and radial frequency to then compare to theoretical predictions.

Exoplanet (planets outside our solar system) atmosphere and ice model simulations: While the codes used for exoplanet simulations does not require HPC, the number and complexity of the runs needed for this project require running on HPC to be completed within reasonable timeframes. I use PlaSIM, a third party code, to investigate ice formation and deposition on tidally locked exoplanets, and then determine if and how this changes the moment of the inertia which may in turn cause the planet to rotate and decrease the likelihood of the exoplanet being habitable.

In our research, we are developing and implementing a multigrid method for the 3-dimensional reference interaction site model (3D-RISM) theory of solvation which is distributed as part of the AmberTools molecular modeling suite. 3D-RISM works by modeling biomolecular systems via their solvent distributions, rather than treating each individual atom separately. Currently, solving the 3D-RISM integral equations is computationally expensive and a method of speeding up these calculations is sought out to reduce the computational costs while maintaining the desired level of accuracy in the calculations. Researchers have shown that a multigrid approach to the 3D-RISM integral equations drastically help to speed up calculations on just a single processor. Our algorithm will be implemented into the 3D-RISM source code, optimized for efficiency, and then parallelized to be compatible with AmberTools’s MPI infrastructure. The results of this work can reasonably be assumed to provide a significant speed-up in calculations, resulting in the ability to model larger biomolecular systems using 3D-RISM.

Intracranial aneurysm rupture accounts for about five percent of all strokes (Stroke accounts for about 1 in 20 deaths in the United States), but the consequences of aneurysmal rupture, and subsequent subarachnoid hemorrhage are dire, carrying a mortality rate of more than 50% and a morbidity rate among survivors of 50%. Successful management of cerebrovascular disease represents an opportunity to improve the outcomes of people suffering from its effects and reduce burgeoning healthcare costs. While improvements in medical imaging constantly increase the accuracy and completeness of  cerebrovascular anatomical information, there still exists a need for quantitative assessment of intracranial aneurysm rupture risk. Hemodynamic modeling can compute mechanical stresses on the arterial wall, but can not predict vessel evolution in the absence of detailed information about the state of the arterial wall. Some preliminary results will be presented on combining modalities for intracranial aneurysms,

Castro is an adaptive mesh refinement (AMR) code, built on the BoxLib library, that solves the compressible hydrodynamics equations for astrophysical flows. The code is parallelized using a hybrid MPI+ tiled OpenMP approach and uses an eulerian grid with simultaneous refinement in space and time. Castro supports a general equation of state and reaction network, full self-gravity and rotation. Gray and multigroup radiation are available under a flux limited diffusion (FLD) approximation based on a mixed frame formulation. We are employing the gray radiation capability to study the black widow pulsar (BWP) binary system. In this system, The fast rotating pulsar emits intense radiation, which injects energy and ablates its low mass companion star. In our setup, we are modeling the companion star with the radiation field as a boundary condition, coming from the bottom of the domain in a 2-d axisymmetry model. We will present the work in progress and future effort. The work at Stony Brook was supported by DOE/Office of Nuclear Physics grant DE-FG02-87ER40317

We present a technique to control the spatial state of a small cloud of interacting bosons at low temperatures with almost perfect fidelity using spatial adiabatic passage. To achieve this, the resonant trap energies of the system are engineered in such a way that a single, well-defined eigenstate connects the initial and desired states and is isolated from the rest of the spectrum. We apply this procedure to the task of separating a well-defined number of particles from an initial cloud and show that it can be implemented in radio-frequency traps using experimentally realistic parameters.

Spectral solar irradiance (SSI), the radiant energy flux per wavelength of the Sun received at Earth, is an important driver of chemical reactions in the Earth’s atmosphere. Accurate measurements of SSI are therefore necessary as an input for global climate models. While models and observations of the spectrally-integrated total solar irradiance (TSI) variations agree within ∼ 95%, they can disagree on the sign and magnitude of the SSI variations. In this work, we examine the contribution of currently-unresolved small-scale magnetic structures to SSI variations in the photosphere. We examine the emergent spectra of two atmospheres with differing imposed-field conditions — one with a small-scale dynamo and the other with a predominantly vertical magnetic field — with similar mean field strengths at wavelengths spanning from visible to infrared. Comparing the radiative output at various viewing angles of pixels of equal vertical magnetic field strength between the two simulations, we find that the small-scale dynamo simulations produce higher radiative output than those in the predominantly vertical field simulation. This implies that the radiative output of a small magnetic structure depends on the magnetic morphology of the environment in which it is embedded, which is currently not included in SSI models.