WHPC Poster Series & Networking Reception

Breaking Barriers and Building Networks: A Hardware Engineer’s Experience with SCINET Fiber at SC25

The SCINET Fiber Team at SC25 offered an invaluable opportunity to contribute to one of the world’s fastest and most sophisticated temporary research networks. Leveraging my existing Level 1 experience in the IEEE networking stack from my work at Purdue’s Rosen Center for Advanced Computing Center, I was promoted to DNOC (Distribution Network Operations Center) Captain, becoming the only female DNOC Captain on the Fiber Team. In the role, I collaborated across technical groups to deploy, verify, and maintain the optical fiber infrastructure that supported thousands of attendees, exhibitors, and HPC researchers throughout the conference.

My responsibilities included installing miles of fiber for more than fourteen exhibitor booths, routing floor fiber across the convention center, and constructing protective thresholds for the floor-run fiber using only masonite and gaffer’s tape to meet safety and reliability requirements. Within the DNOC, I performed cabling work, documented routing paths, and participated in hands-on fiber diagnostics and splicing demonstrations. Daily Tasks also included conducting light-level tests, validating link integrity, and assisting with help desk tickets to ensure exhibitors continued to have uninterrupted connectivity during peak conference activity.

As a DNOC Captain, I also held leadership and mentoring responsibilities. I supported volunteers, including students with varying levels of experience by guiding them through proper installation techniques, safety practices, and fiber handling procedures. This mentoring role strengthened team efficiency while fostering a more inclusive pipeline of future HPC and network engineers.

My experience at SC25 directly translates back to Purdue’s data center operations, where the skills gained have already improved my ability to diagnose connectivity issues rapidly, optimize cable management during cluster deployments, and refine troubleshooting workflows. Serving on the Fiber Team not only advanced my technical capabilities but also enhanced the visibility of women in HPC hardware engineering, reinforcing the importance of diverse leadership within large-scale computing environments.

AI-based Catalytic Performance Prediction for CO2 Electrochemical Reduction Using Ionic Liquids

Recently, ionic liquids (ILs) have garnered remarkable attention as electrolytes for CO2 electrochemical reduction (CO2ER) due to their unique properties viz. thermal and chemical stability, good CO2 solubility, and their potential to reduce overpotential. The catalytic performance of ILs in CO2ER has been explored via experimental methods which have limitations, given the unclear understanding of the complex reaction mechanisms. Recently, Artificial Intelligence (AI) methods have gained increased attention across diverse applications including chemical engineering. These methods play a pivotal role in extracting insights, understanding patterns, and mitigating uncertainty within datasets. In this study, we investigate two categories of AI models (ML and DL) for predicting the CO2ER catalytic performance via Gibbs free energy and capacity. For this task, we formulate a novel dataset (CO2ERIL) from 90 ILs in two formats using: (1) 30 electronic and geometric IL properties, (2) IL chemical structures in SMILES data format. The dataset versions are formulated using the Conductor-like Screening Model for Realistic Solvents (COSMO-RS) and TURBOMOLE software. The prediction results of the AI models on the two dataset versions depict a similar trend in the predicted target variables signifying that the dataset format does not really affect the model performance. Moreover, the best ML model outperforms the DL model for one target variable, Free energy predictions, but lags in the other target variable, capacity predictions. Our study is novel in terms of AI methodology as well as the CO2ER dataset. Our research outcomes will contribute to the more effective selection of ILs for CO2ER catalysis.

Hierarchical Precision and Recursion for Accelerating Symmetric Linear Solves on MXUs

Symmetric linear solves are fundamental to a wide range of scientific and engineering applications, from climate modeling and structural analysis to machine learning and optimization. These workloads often rely on Cholesky (POTRF) decomposition and its supporting operations—triangular solves (TRSM) and symmetric rank-k updates (SYRK)—which together form the computational core for solving symmetric positive definite systems. To accelerate these kernels, we present a portable, mixed-precision solver designed for Matrix Processing Units (MXUs), including NVIDIA Tensor Cores (H200) and AMD Matrix Cores (MI300X). Our algorithm builds on a nested recursive formulation in which Cholesky exposes parallelism through recursive decomposition of its TRSM and SYRK subproblems, incorporating the first recursive GPU implementation of SYRK. This structure yields a hierarchical recursion that maximizes GEMM throughput while enabling fine-grained control over numerical precision. We introduce a custom recursive data structure that assigns low-precision FP16 arithmetic to large off-diagonal blocks, while preserving high precision on diagonal blocks to ensure numerical stability. To mitigate the limited dynamic range of FP16, we integrate a lightweight block-wise quantization scheme that prevents numerical overflow.

The solver is implemented in Julia, leveraging array programming, multiple dispatch, and dynamic type inference to enable seamless expression of mixed-precision computation. This design provides a high-level, hardware-agnostic interface while efficiently interfacing with low-level vendor libraries for backend portability. On H200, our recursive FP64 SYRK achieves a 14× speedup over cuBLAS, while mixed-precision delivers up to 27.0× speedup in SYRK and 5.3× in TRSM over full-precision baselines. This results in a 5.32× overall speedup for Cholesky versus cuSOLVER FP64, with 100× better accuracy than pure FP16 while retaining 88% of its peak speedup. Comparable performance and accuracy trends are observed on MI300X, demonstrating broad applicability across GPUs.

Accelerating Multi-Omics Analysis with High-Performance Computing

High performance computing (HPC) has become essential for analysing increasingly large and complex multi omics datasets. Modern biomedical research frequently integrates diverse molecular layers—including genomics, epigenomics, proteomics, metabolomics, and phenomics—to achieve a more comprehensive understanding of disease mechanisms. However, the scale, heterogeneity, and computational intensity of such analyses exceed the capabilities of standard desktop environments. This work demonstrates how HPC infrastructures accelerate multi omics workflows, improve computational stability, and support reproducible large scale analyses using data from the Northern Ireland Cohort for the Longitudinal Study of Ageing (NICOLA).

Whole Genome Sequencing (WGS) presents one of the most computationally demanding components of multi omics pipelines. The processing of WGS data requires sequential steps involving quality control, alignment to a reference genome, variant calling, and annotation. Using the FastQC module and the nf-core/sarek pipeline on an HPC cluster, we were able to efficiently process ~1 TB of sequencing data across 146 samples. The parallelisation and memory optimisation achievable on HPC reduced turnaround times substantially. Workflows that typically require multiple days on a standard desktop—often resulting in system crashes or memory failures—were completed within 24 hours on HPC. This demonstrates the critical role of HPC in delivering scalable, reliable WGS analysis for medium to large cohorts.

Beyond sequencing, multi omics integration poses its own computational challenges. For the NICOLA cohort, we incorporated several datasets of varying sizes and complexity: epigenetic profiles (~900,000 CpG sites, ~2 GB), Biocrates metabolomics (~1019 metabolites), Nightingale NMR metabolomics (~247 metabolites), and OLINK proteomics (~5000 proteins, ~28 MB). We employed Multi Omics Factor Analysis (MOFA) to uncover shared latent structures and cross platform molecular signatures. MOFA’s computational demands scale rapidly with increasing sample size and feature numbers, often exceeding the memory capacity of typical desktop machines. Running MOFA on the HPC cluster reduced runtime from 4–6 hours to approximately 3 minutes and eliminated common issues such as crashes during matrix factorisation, highlighting the importance of HPC in enabling stable high dimensional modelling.

Overall, HPC provided four major advantages for this research: (1) accelerated analysis, enabling rapid turnaround for both sequencing and integrative modelling; (2) stability, preventing crashes during memory intensive steps; (3) reproducibility, supported through structured job scripts and shared environment modules; and (4) efficient resource utilisation, ensuring optimal balancing of CPU and memory loads across the cluster. Together, these benefits significantly enhance the feasibility, robustness, and scalability of advanced multi omics epidemiological analyses.

This work illustrates how HPC transforms the analytical capacity of biomedical research by overcoming computational bottlenecks and enabling the integration of diverse biomolecular datasets. As multi omics continues to expand in complexity and scale, HPC will remain an indispensable tool for unlocking deeper biological insights and supporting reproducible, large cohort molecular epidemiology.

Co-Scheduling – Towards Energy Efficient HPC Operations

With the growing usage of high-performance computing (HPC) across multiple application domains, energy consumption has become a topic impossible to ignore. Among various energy saving mechanisms, oversubscription: allocating more computational tasks than the physical capacity of the system, has emerged as a suitable approach. In modern HPC nodes, which contain many cores with shared memory channels and L3 cache, memory bandwidth and cache contention are key performance and energy bottlenecks.

HPC workloads can broadly be classified into two. Memory bound applications stress memory bandwidth and exhibit high cache miss rates, consuming significant energy for data retrieval. In contrast, compute bound applications primarily stress CPU cores, have a high compute-to-memory ratio, and spend most of their energy on computation. These complementary resource demands create an opportunity to run such workloads simultaneously, sharing system resources more efficiently through oversubscription.

This work investigates oversubscription through co-scheduling of HPC workloads. All experiments were conducted on the HSUper HPC system, which consists of two sockets with 36 cores each (72 physical cores total), using SLURM as the scheduler. As an initial test case, we used Stream, a memory bound benchmark, and Linpack, a compute bound benchmark. The goal was to identify an effective scheduling configuration for oversubscription. Both applications were pinned to specific cores on a single node using likwid-pin.

Multiple configurations were evaluated, ranging from fully overlapping core assignments to completely non-overlapping ones. These configurations also varied in how memory channels were saturated. In total, eight scenarios were tested, and performance and energy consumption were measured. Results showed that the optimal configuration involved running the two applications on non-overlapping cores while saturating all memory channels. In this scenario, Stream experienced a performance degradation of 3.13% and Linpack 26.15%, while overall energy consumption was reduced by 3.7%. Although the cores were non-overlapping, the L3 cache and memory bandwidth were still shared, making this approach, co-scheduling, even though it’s not pure oversubscription.

After identifying a suitable configuration, the co-scheduling study was extended to real applications. In the next experiment, Stream was replaced by Lesocc (Large Eddy Simulation on Curvilinear Co-ordinates), a 3D finite volume code for computational fluid dynamics on block-structured grids with non-orthogonal and non-staggered arbitrary geometry. Lesocc was co-scheduled alongside Linpack. The results showed that co-scheduling reduced the total runtime by 10 minutes and lead to a reduction of 20% in energy consumption. As expected, performance degradation was observed in both applications caused by resource contention. However, the reduction was within a tolerable range considering the substantial energy savings achieved.

Overall, these results demonstrate that co-scheduling complementary workloads can significantly improve the energy efficiency of HPC systems, at the cost of moderate performance degradation. Ongoing work includes deeper analysis of L3 cache behaviour and extending to other compute bound applications. Ultimately in future, we aim at generalizing the applicability and providing a co-scheduling configuration to HSUper users.

High-Performance Computing Workflow for Rapid-Response Physics-Based 3D Earthquake Simulations Integrated In Data Lake

Dynamic rupture simulations are the gold standard in physics-based earthquake simulations and combine modeling seismic wave propagation with non-linear frictional failure of subsurface fault systems. Despite efficient scaling on Tier-0 supercomputers, setting up and running full workflows with high-resolution simulations after a major earthquake can take weeks to months. We introduce a rapid-response workflow that leverages a precomputed scenario catalog generated with SeisSol (500,000 core-hours on SuperMUC-NG@LRZ) for Alto Tiberina, Italy. Our workflow could be triggered when a major earthquake strikes; it rapidly queries the SeisSol catalog by comparing observations with precomputed synthetic seismograms. This process identifies the most representative scenario and provides precomputed shake maps, accelerating this DT-Geo “digital twin component” to less than 15 minutes.

The research data management of extensive parameter studies is complex, and the deployment of efficient storage solutions is increasingly important to avoid losing expensive research datasets.
Several HPC Centers and Research Infrastructures are investing in storage solutions for simulation data. We integrated this workflow into CINECA’s new Simulation Data Lake (SDL), offered through Geo-INQUIRE to Earth scientists in Europe. The SDL relies on a comprehensive metadata scheme that enables the storage of a wide variety of data and is already prepared to be integrated within the EPOS data portal. We provide a set of typical SeisSol files to benchmark data up- and downloading to the SDL. All catalog data, synthetic seismograms, and corresponding shake maps are shared FAIR-compliant through the SDL and will be accessible to CINECA’s AI Factory.

Our open-source workflow, data, and code are freely available and can be easily adapted to additional regions or alternative earthquake simulation software.

Training Multi-Billion-Parameter AI Weather Models with Optimized Domain and Tensor Parallelism

AI-based methods have rapidly revolutionized atmospheric modeling, with recent successes in medium-range forecasting spurring the development of foundation models. However, accurate modeling of complex atmospheric dynamics at high spatial resolutions requires billions of neural network parameters and gigabyte-sized data samples, making accelerator memory and I/O-bandwidth the bottlenecks for model training. To overcome these limitations, we introduce Jigsaw, a distributed training and inference scheme that leverages domain and tensor parallelism to eliminate memory redundancy across model-parallel processes and reduce I/O demands. We apply the Jigsaw parallelization scheme into an MLP-Mixer architecture, WeatherMixer, a multi-layer-perceptron-based model with global vision that is well-suited for learning weather phenomena. Using Jigsaw, we train WeatherMixer with up to 3.2B-parameters, achieving predictive performance competitive with numerical weather prediction and state-of-the-art AI models. To highlight the computational performance, we perform scaling experiments on global 0.25° (~30 km resolution) ERA5 data across two HPC systems. Anticipating that future reanalysis datasets will include even higher resolutions, we demonstrate, for the first time, training on 0.125° data.

The scaling experiments demonstrate that high-resolution input data samples benefit from domain parallelism and improve per-GPU computational throughput by reducing dataloading bottlenecks. In compute–communication–limited regimes, Jigsaw achieves state-of-the-art performance in distributed model training, with 97% of theoretical peak performance on 4 GPUs; and a strong scaling speedup of 6.4 when training across 8 GPUs. By combining domain, tensor, and data parallelism at larger scales, training on 256 GPUs reaches 11 PFLOPs with a scaling efficiency of 72% compared to 51% without Jigsaw.

Advancing The Data Assimilation Research Testbed (DART) as an Early-Career Software Engineer

The Data Assimilation Research Testbed, or DART, is an open-source, freely available software facility for ensemble data assimilation, supporting a vast and diverse community of users for over 20 years. Ensemble data assimilation is a statistical method that is used to combine information from numerical model predictions with measurements of the Earth system to enhance the value of both. Applications include ensemble forecasting, generating initial conditions for forecasts and predictability studies, diagnosing model error and bias, and assessing the value of existing and planned observations.

DART includes interfaces to many models and observations across all Earth system domains. DART also allows for the use of a variety of assimilation algorithms, including novel methods developed by the DAReS team that are especially effective for pollutants, sea ice concentration, and soil moisture. Our modular software design and generalized model interface routines facilitates the addition of new models, observations, and assimilation algorithms and allows for users to easily switch between the available options.

Focusing on accessibility, our software is carefully engineered to run efficiently on systems ranging from laptops to top performing supercomputers. DART utilizes the power of HPC through distributed memory and parallel computation with the Message Passing Interface (MPI) to be able to run with complex, high resolution models and large numbers of observations and ensemble members. This further extends the versatility of DART by allowing users to work on more computationally demanding projects.

For these reasons, DART is able to support a diverse, long-standing, and continuously growing user community. This inherent diversity in DART has also enabled me to contribute to a broad spectrum of projects, each requiring unique skills and technologies.

This poster will detail the versatility of DART and its functionalities, showing how this has shaped my work and early-career experiences as one of the two software engineers on the team. Working with DART has allowed me to engage with a wide variety of systems, technologies, programming languages, assimilation algorithms, models, observations, collaborators, and background science.

The poster will also describe my varied technical contributions and specifically highlight a few of my larger projects that showcase the interdisciplinary and multifaceted nature of my role in DAReS. Examples include collaborative projects with the NASA Goddard Space Flight Center and the NSF NCAR High Altitude Observatory (HAO), facilitating the use of DART data assimilation with space weather models on NASA’s Pleiades Supercomputer and the co-development of pyDARTdiags, a Python package for manipulating observation sequences and calculating observation-space diagnostics for DART.

From Prototype to Production: Agentic AI Tools for Natural Language-Driven HPC Management

High-performance computing (HPC) and AI infrastructure are growing rapidly, placing increasing operational pressure on HPC service providers. In the UK, the deployment of Isambard AI, one of the world’s fastest supercomputers, supports hundreds of projects and thousands of users, with frequent AI research calls and dynamic compute allocations. Managing access, allocations, and usage reporting at this scale presents significant challenges. While platforms such as Waldur manage users, projects, and compute resources, many reporting and administrative workflows remain manual, particularly when responding to external stakeholders such as UKRI and DSIT who require periodic usage and consumption reports. These processes are time-consuming, error-prone, and limit real-time visibility for both HPC staff and users.

This work explores how modern AI protocols can be used to automate and simplify HPC operations. We integrate large language models with live infrastructure data using the Model Context Protocol (MCP), a standardised client-server framework that enables AI assistants to securely query APIs, retrieve structured data, and execute controlled actions. By exposing Waldur’s APIs as MCP tools, users can issue natural-language queries to retrieve real-time project statistics, usage summaries, and administrative information, while complex requests are automatically decomposed into multiple tool calls with structured orchestration.

A key contribution of this work is the secure transition from prototype to production. We implement OpenID Connect (OIDC) device authentication using Keycloak to support MCP clients operating in chat-based environments where traditional browser redirect flows are impractical. Users authenticate via a browser-based device flow, after which tokens are securely exchanged and introspected before issuing Waldur API tokens. Read-only and read-write MCP servers are strictly separated, and all write operations are permission-checked to prevent unauthorised or “silent” actions. No credentials are hard-coded, and each user authenticates independently.

The resulting system enables automated reporting, real-time dashboards, and natural-language interaction with HPC management platforms, significantly reducing manual workload and improving operational visibility. Beyond raw data retrieval, the system generates narrative summaries, trends, dashboards, and recommendations to support decision-making, usage forecasting, and resource optimisation. This work demonstrates how AI combined with modern protocols such as MCP and OIDC can form a secure, scalable interface for managing complex HPC infrastructures.

Learning Gaussian-Like Deposition to Reduce PIC Noise in HEXAPIC at Scale

Particle-in-Cell (PIC) is a first-principles approach for simulating collisionless or weakly collisional plasmas. It advances charged macroparticles while computing electromagnetic fields on a grid, enabling kinetic modeling for applications such as fusion edge physics, plasma-based materials processing, electric propulsion, and space/astrophysical plasmas. Despite its importance, PIC remains expensive at scale. In practice, a PIC study is not only a timestep loop; it is an end-to-end workflow that includes data generation, preprocessing, large production runs, diagnostics, and repeated parameter scans. At scale, performance is frequently constrained by memory bandwidth, data movement, and interconnect communication, not only by floating-point throughput.

We present this work in the context of HEXAPIC (Heterogeneous Exascale PIC), an exascale-oriented PIC code designed for heterogeneous HPC systems. HEXAPIC provides scalable infrastructure for large kinetic simulations, but a persistent limitation is numerical noise from finite macroparticle sampling, especially in the particle–grid coupling step. The commonly used cloud-in-cell (CIC) deposition/interpolation is efficient and scalable, yet it can introduce grid-scale fluctuations that obscure weak physical signals and degrade diagnostic quality. Moving from CIC to higher-order (smoother) particle shape functions can reduce this noise, but typically increases computational work and, more importantly for HPC, increases memory traffic and can reduce locality, key concerns on bandwidth-limited architectures.

To address this trade-off, we investigate an ML-assisted approach designed to fit the HEXAPIC execution model. Rather than performing ML inference per particle, we use a PIC-friendly two-step strategy: (1) deposit with CIC, then (2) apply a lightweight learned smoothing operator on the grid that maps CIC-deposited fields to a Gaussian-like deposition. The surrogate is represented as a small convolution kernel (separable or full 2D) with non-negativity and sum-to-one constraints to approximately preserve total charge (up to boundary effects). A major focus is practical HPC integration: minimizing host-device transfers, bounding memory footprint, preserving locality under domain decomposition, and avoiding additional synchronization or communication amplification. Training and experimentation use PyTorch parallelism (data parallelism, and model parallelism where relevant) with a current baseline on multicore CPUs and a single GPU.

Preliminary training on large PIC-derived datasets shows stable optimization and consistent generalization trends for the Gaussian-structured regression model, motivating the next phase of physics- and numerics-driven evaluation. Planned validation includes variance reduction and attenuation of high-wavenumber spectral content in deposited quantities, alongside checks that the learned smoothing does not introduce unphysical artifacts in downstream fields and diagnostics. Overall, this work demonstrates a practical route to ML-based noise reduction in HEXAPIC that targets improved diagnostic quality while maintaining scalability on heterogeneous exascale systems.

A case study in sustainable AI: Applying carbon-reducing techniques in weather forecasting frameworks

Machine Learning (ML) is transforming weather and climate science, increasingly complementing traditional methodologies for forecasting and analysing complex atmospheric phenomena. There is evidence that ML-based solutions can be computationally more efficient than traditional models while achieving comparable accuracy. As minimising the energy consumed by data centres and HPC resources is key to help tackle climate change, leveraging ML for scientific computing applications has the potential to significantly reduce the environmental impact of weather and climate modelling.
We are investigating how a broad set of sustainable AI techniques across data curation, model selection, training and inference can be applied when using domain-specific ML frameworks. This project explores to what extent sustainability is already built-in, and whether the abstractions that these frameworks provide aid or hinder the application of sustainable AI best practices. Driven by the observation that accessibility of these techniques is a blocker to adoption, as considerable expertise and effort can be required to leverage ML within scientific workflows, domain-specific ML frameworks have become popular recently. Examples include ECMWF’s Anemoi and NSF NCAR’s CREDIT, and these have demonstrated the potential to significantly lower the barrier to undertaking research into data-driven weather and climate prediction. These technologies accelerate science by providing tools that abstract away many of the time-consuming details that form part of ML research into configuration files.
Whilst ML models have the potential to offer a step change in simulation efficiency, undertaking ML research can incur significant environmental impacts. There are some potential mitigations, for instance the carbon execution cost can be decreased through green scheduling, and optimising ML workflows to improve both model accuracy and efficiency has been demonstrated to be an effective strategy. However, these efforts can depend very much on location (e.g. low carbon energy). A key objective of this project is to collate appropriate techniques across the community and document their potential impact.

Identifying appropriate metric(s) is a key consideration when exploring the sustainability benefits and challenges of fusing ML with scientific computing. Common approaches include reporting the runtime of a model’s final training run, or the amount of energy or water consumed by one inference run. However, such measurements can be myopic and it is important to consider energy-efficiency and more broadly sustainability for the entire development lifecycle: From data acquisition, preparation, the effective use of resources during (re-)training and tuning hyperparameters, to the integration of a trained model into existing systems and workflows.

This work aims to better understand current best practice, limitations and opportunities around fusing ML with scientific workflows to drive energy efficiency. It is part of the CONTINENTS project, a four-year long EPSRC-funded collaborative programme between EPCC at the University of Edinburgh, the UK’s National Centre for Atmospheric Science (NCAS),and the US National Center for Atmospheric Research (NCAR). This interdisciplinary collaboration aims to bring together world leading centres of HPC research and service provision, atmospheric science experts, and numerical and ML application developers to provide a step change in sustainable climate and weather modelling.