Student Posters Repository

Software tools are integral components of the HPC software stack and provide invaluable measurements and insights into application run time and system behaviour to end users, code developers and system administrators. However, most tools currently do not support performance analysis at the granularity of libraries, which are the most important level of abstraction for code when developing modern applications. To overcome this limitation, we present a novel infrastructure that can auto-generate tool interfaces that enable interception at library-level. This opens the door to deploying tools at the right level of abstraction and with that to many use cases previous impossible or infeasible. 

Massive stars end their lives in supernovae, powerful explosions that rank among the most energetic events in the universe. These explosions release approximately 10^58 neutrinos, which carry away 99% of the gravitational binding energy. The resulting compact objects may eventually merge, emitting material rich in heavy elements into the galactic medium. Neutrinos play a fundamental role in these explosive stellar events. In supernovae, they transport the energy that revives the stalled shock wave, leading to the explosion. In binary mergers, neutrinos control the numbers of protons and neutrons available for nuclear processes that generate heavy elements in the universe. However, neutrino transport presents significant computational challenges and is currently one of the largest sources of uncertainty in simulations. Addressing this issue, I present recent advancements in EMU, a particle-in-cell based, quantum neutrino transport code built under the AMReX framework, which is performance-portable for both GPU and CPU. I discuss key computational aspects of the code and the neutrino flavor dynamics observed in high-resolution simulations of dense neutrino gases under realistic conditions in core-collapse supernovae and neutron star mergers, including the appearance of fast and collisional flavor instabilities and chaos.

The presence of organic acids in fast pyrolysis bio-oil significantly impacts its suitability for various fuel applications, with acetic acid constituting up to 15% of the bio-oil. Esterification, a promising approach to reduce acid content by forming esters, can be economically facilitated using crude glycerol, an underutilized by-product from the biodiesel industry. This study systematically investigated autocatalytic esterification of glycerol with acetic acid at varying mole ratios and temperatures. Employing 2D-NMR and GC-MS, various glycerol esters were identified and quantified for the first time in the field. Maximum glycerol conversion reached 26.25%, 59.08%, and 95.62% at 90°C, 120°C, and 150°C, respectively, with 1-monoacylglycerol being the predominant product. Computationally, Condensed Dual Descriptors were used to quantitate reactivity of each hydroxyl group revealing insightful differences to electrophilic susceptibility between the middle and terminal hydroxyl groups. Furthermore, symmetry adapted perturbation theory techniques (SAPT) were then used to compare  the interaction strength between reactant complexes of different isomers, revealing DFT simulations provided insights into the esterification mechanism, revealing lower energy barriers for 1-monoacylglycerol. The study presents a comprehensive understanding of the autocatalytic esterification mechanism and lays the groundwork for enhancing fast pyrolysis bio-oil quality through crude glycerol esterification.

General-purpose processors are sensitive to Single Event Upsets (SEUs), a type of transient fault flipping a bit in SRAM, caused by e.g. cosmic rays. The effects of SEUs are particularly noticeable in HPC applications, as the need for a large number of CPU hours virtually guarantees such an event occurs. Software-based fault-tolerance allows the use of commodity processors, but relies on replication. The resulting overhead which may be too significant, while perfect protection need not be required as algorithms may tolerate some imprecision, e.g. those based on iterative optimization. Combining both the need for fault-tolerance and the intrinsic ability to tolerate some faults, we present "Strategy Switching". A strategy is a partial application of fault-tolerance to the application code, offering a balance between reliability and resource usage. By changing strategies at runtime, we further improve the reliability of the system by ensuring no part of the application is disproportionately affected by faults. Consequently, we maintain the quality of the output while saving resources. We are developing Strategy Switching for multi-node and multi-cluster HPC applications, necessitating accurate modeling of resource allocation, transfer times and network congestion. We have already demonstrated the efficacy of Strategy Switching for single-node real-time applications.

For the European Commission's Destination Earth project, we at the Barcelona Supercomputing Center are involved in the development of digital twins of the earth. The ambitious goal of the project is to create high-resolution digital twins of the earth at an unprecedented sub-km scale. In order to achieve these goals, we have to accelerate the computation of a model by porting regions to GPUs. These regions are found through profiling, by using AMD's rocprof, NVIDIA Insight, and the compiler (runtime) output. We port the regions to GPU using OpenACC, which has been chosen due to the time constraints of the project. The difficulty of attaining high performance is worsened due to the requirement of supporting both the AMD and NVIDIA platforms. For this project, we need to support multiple EuroHPC clusters, such as MareNostrum 5, Lumi, and Leonardo. The increased difficulty is due to each platform having its own hardware and software stack, influencing the efficiency of the OpenACC port.

I am working to understand the coevolution of galaxies and astrophysical dust through the use of high-resolution numerical simulations. Astrophysical dust (or composite solids made of conglomerations of individual atoms) is known to be ubiquitous in galaxies and is one of our main proxies for observing them. However, much remains to be learned about how dust forms, evolves, and is distributed throughout galaxies. My research involves creating high-resolution 3D hydrodynamical simulations of dusty galaxies, allowing us to watch the processes that govern dust’s evolution as they unfold. I am contributing to the Cholla (Computational Hydrodynamics on paraLLel Architectures) code to simulate these galaxies at field-leading resolution. Cholla is a GPU-native code that has been designed to run on the largest supercomputers in the world. Cholla utilizes CUDA/HIP to calculate fluid properties of cells in a Cartesian grid on individual cores of GPUs and can be expanded to run with even larger simulation volumes on many GPUs (with nearly perfect weak scaling) through the use of MPI. Through its participation in the Frontier Center for Accelerated Application Readiness (CAAR) program, Cholla has been optimized to run simulations on the Oak Ridge Leadership Computing Facility’s exascale supercomputer, Frontier.

Exascale systems are now appearing in the US, Japan, in Europe with two exascale procurements in the coming years and very likely in China. These systems offer great opportunities for the numerical astrophysics community and also raise serious technical challenges.  Many legacy codes need to be fully rewritten to benefit from exascale architectures.  Performance portability is a key issue as well as the management of the huge amount of data generated by exascale simulations.

In this presentation I will discuss the path we have chosen to exascale and the software engineering solution chosen to address performance portability in the long run. I will present a new radiation hydrodynamics code based on the Kokkos library. The first target of this new code will be to study supernova ejecta. They are multidimensional fluids, structured on both small and large scales following the non-linear development of complex fluid instabilities during and after the explosion. They contain a wealth of information about stellar composition and the evolution of elements in the interstellar medium. Their study requires significant computing resources that can be met through the use of exascale machines and suitable software. 

Solar flares cannot be explained or simulated without considering the effects of accelerated particles. State of the art magnetohydrodynamic (fluid plasma/MHD) simulations are able to reproduce large scales and slow phenomena in the solar atmosphere, but not solar flare kernels. Why? Because flare kernels do not behave like fluids. 

Satellite observations of solar flares indicate electrons and ions at relativistic speeds. Such speeds break any fluid assumption. We need kinetic simulations, but they are yet too expensive due to the massive scale separation between flare kernels and the flare exterior, which play an important part.

A stepping stone is to understand the underlying particle acceleration mechanisms, which are still heavily debated. I my project, we seek to reveal the acceleration mechanisms through hybrid MHD/test-particle (Monte Carlo) simulations.

Interacting systems like the core collapse supernovae permit many-body quantum dynamics. These systems develop entanglement entropy through many-body neutrino correlations because of the intrinsic property of neutrino spins to oscillate between its different flavors. My research models this quantum physics in supernovae via tensor networks to understand the impact of neutrino’s entanglement on the collapsing star. For this project, I attempted to parallelize the most crucial parts from my main code, which were identified as the main bottlenecks, utilizing multithreading in two ways to gain computational efficiency. I then include the results and analysis I made from each parallelization.