Student Posters Repository

Simulating turbulence in magnetic confinement fusion devices is a non-trivial task that requires a large number of computing resources. The recently developed GENE-X extends the typical coverage of turbulence simulations to the edge and scrape-off layer of magnetic confinement fusion devices, making it highly resource-demanding. Currently, GENE-X uses a heterogenous parallelization featuring OpenMP for intranode and MPI for internode parallelism respectively. GPU offloading via Fortran/C++ interface is being implemented to enable simulations on broader range of devices and physics models. Fortran/C++ interfacing approach is chosen for more robust GPU offloading backend and support. GENE-X also has many application-specific numerical and platform-level parameters which the user needs to configure manually and it is suboptimal. Inspired by autotuning in deep learning applications, ML-based autotuning is considered to be the solution since it can avoid exhaustive search in combinatorial parameter space. More cost-effective search can be done with ML-based autotuning via Bayesian optimization as currently being considered.

The field of research I am working on is radiative heat transfer within materials. We develop simulation softwares that can compute the material response of objects placed under extreme heat loads like the heat shields (Thermal protection systems) of re-entry bodies. Amongst the physical phenomena involved in such an environment, radiative heat transfer is a significant contributor to the material response. Modelling the radiative intensity field is more complicated than the typical scalar or vector fields as radiative intensity depends on direction and spectrum in addition to space and time. We use special models in the directional and spectral space to simplify the modelling of radiative intensity.

Hardware heterogeneity in HPC is increasing at the dawn of the exascale era, with performance portable APIs such as Kokkos rising to meet the challenge. Accompanying the increase in heterogeneity is an increase in transient faults and soft errors, and industry leaders have identified software resiliency as a critical research need. Resilient Kokkos proposes a framework to tolerate soft errors during execution based on a primary Kokkos abstraction, the execution space. The resilient execution space allows applications already using Kokkos to take advantage of a near-seamless API integration to obtain a user-specified level of resiliency in parallel execution. Soft error detection and mitigation is handled by a majority vote in resilient versions of parallel execution abstractions. We show performance results on various benchmarks written for Kokkkos and modified to use the resilient execution space.

Frustrated magnets are a platform to explore new phases of matter. This includes the quantum spin liquid, a magnetically disordered phase carrying fractional excitations. One method to generate frustration is by including bond-dependent dipole interactions such as in the Kitaev honeycomb spin-1/2 model, which is exactly-solvable and hosts the Kitaev spin liquid (KSL). Moreover, these interactions appear naturally in effective pseudospin models of spin-orbit coupled Mott insulators. In some insulators, however, the low-energy degrees of freedom lack a dipole moment but feature bond-dependent quadrupole and octupole interactions; a natural question is whether such a system could host a disordered phase analogous to the KSL. We find that an exactly-solvable Kitaev multipolar liquid can be stabilized due to quadrupole-octupole interactions induced by an out-of-plane magnetic field. We then use several numerical methods to map out the classical and quantum phase diagrams, including classical Monte Carlo and exact diagonalization respectively. The use of HPC protocols such as MPI, OpenMP, and GNU Parallel in these simulations is discussed.

In building equipment for space exploitation, one has to trade system robustness for the high processing capabilities and low energy consumption. The high performance, low robustness approach, is acceptable, especially in Earth’s vicinity. However, in more demanding (especially heavily radiative) environments, attempts had disappointing outcomes. The processing-reliability gap, between highly reliable and highly performant systems, spans 2-3 orders of magnitude (Fig. 1). This gap brings hope, that some of this excess processing power can be utilized, in building a combination of hardware and software mechanisms that is capable of increasing the robustness and resilience of otherwise susceptible semiconductor devices, while allowing to harness the remaining processing power to build affordable space systems with large degrees of autonomy, rich functionality, and high bandwidth. At the CritiX research group, we aim to bridge this performance-reliability gap, by researching the enabling building blocks for constructing more reliable and more secure System-on-Chips.