Abstract: Ubiquitous in-node heterogeneity intensifies the challenges of programming productivity, portability, and concurrent utilization of different co-existing architectures. Developed at Oak Ridge National Laboratory, the 2024 R&D Award winner, IRIS Runtime is an intelligent task-based runtime system designed to alleviate such challenges for diverse heterogeneous systems. Inspired by the philosophy of "implement once and deploy anywhere," IRIS provides a unique programming model and runtime environment that enables porting IRIS applications written in high-level languages ( C/C++, Fortran, Python, Julia, etc.) to different heterogeneous systems without changing the source. While providing portability, IRIS facilitates seamless concurrent execution on mainstream vendors' CPUs, GPUs, FPGAs, and DSPs. Serial and architecture-agnostic IRIS programs do not need any data transfer specification where the runtime identifies concurrency, schedules the computation, automatically and optimally moves data among devices (outstanding paper at HPEC'23), and ensures computation and communication overlap through asynchronous execution by using vendor and open-source APIs for different architectures. The ORNL IRIS team, in collaboration with researchers from Carnegie Mellon University, US, and the University of Tsukuba, Japan, is building a set of IRIS abstractions that have the potential to expose various auto-tuning opportunities.
Biography: Mohammad Alaul Haque Monil is a research scientist in the Computer Science and Mathematics Division at Oak Ridge National Laboratory. Monil's research interests include heterogeneous runtime systems, math libraries, performance modeling, measurement, and analysis of heterogeneous systems. He is one of the leading developers and researchers of the 2024 R&D 100 Award winner, IRIS-SDK, where his notable contributions include math library-MatRIS, heterogeneous memory management, scheduling, etc. Monil is an early career researcher who published 30 conference and journal papers in various venues; conferences include PACT, IPDPS, ICPP, ISC, HPEC (outstanding paper at HPEC'23), etc. Monil is an ACM member. He received his Ph.D. in Computer Science from the University of Oregon in 2021; he also received his MS and BSc degrees from North South University and Khulna University of Engineering and Technology, respectively.

Abstract: This talk offers a data-centric perspective on the future of HPC performance engineering, emphasizing reproducibility, automation, and transparency. We explore recent advances in analyzing and modeling parallel I/O behavior, and discuss tool-agnostic workflows that support scalable and explainable performance studies. Highlighting research efforts such as MAWA-HPC, I/O roofline modeling, JUBE-ML, and VerifyIO, we illustrate how principled data collection and reproducibility-aware methodologies can help bridge the gap between observation and optimization in complex HPC environments. Our aim is to foster trust in performance insights and promote sustainable practices in high-performance systems research.
Biography: Sarah Neuwirth is a Full Professor for Computer Science and Chair of the "High Performance Computing and its Applications" research group at Johannes Gutenberg University Mainz (JGU). She serves as the Co-Director of the NHR South-West HPC Center. In 2018, Sarah completed her PhD in computer science at Heidelberg University. Her research interests include parallel I/O and storage systems, modular supercomputing, performance modeling and analysis, optimization, reproducible benchmarking, and parallel programming models. For her outstanding contributions to HPC, Sarah was awarded the "2023 PRACE Ada Lovelace Award for HPC" and the "ZONTA Science Award 2019". She has participated in numerous research collaborations as co-PI, including working with: Jülich Supercomputing Centre (DEEP Project Series, EUPEX), LLNL, BITS Pilani Goa Campus, ORNL, and Virginia Tech.

Abstract: Automatic tuning of control algorithms is key to effectively track set points and to reject disturbances in machine control and robotics applications. As such, a key question in adaptive control is to elucidate algorithms to automatically tune the behavior of control algorithms to accommodate diverse control problems/tasks with minimal intervention and under tight computational budgets. In this talk, we describe our recent findings from a new class of algorithm extending differential evolution with particle adaptation schemes and the learning of activation functions from neural inference. Computational experiments considering five machine control scenarios including motor position control, motor velocity control, crane stabilization, inverted pendulum, and magnetic levitation demonstrate the superior control and generalization performance compared to relevant existing frameworks. Our results have the potential to further advance towards developing efficient and self-adaptive tuning algorithms which may find use in a wider set of optimization and control problems.
Biography: Victor Parque is an Associate Professor at Hiroshima University, Japan, his interests span the principles and applications of computational learning and evolutionary systems in mechatronics, robotics, computing, optimization, design engineering, planning, and control. He is the author of more than 100 peer-reviewed papers in journals and international conferences of well-known academic societies. He has been honored with several competitive research grants and is actively involved in collaboration with the industry. He was honored as a finalist in the Hummies Awards for Human-Competitive Results from the ACM Genetic and Evolutionary Computation Conference, an Honorary Diploma in Systems Engineering, the Outstanding Paper Award from CANDAR, and the distinction of top 10% selected papers at the International Conference on Engineering Design.

Abstract: AI development has become increasingly driven by powerful frameworks like PyTorch and TensorFlow, supported by major tech companies. However, the rapid release cycles of these frameworks -- every 3-6 months -- pose a challenge for new hardware vendors. They struggle to develop the necessary AI functionality and keep pace with frequent updates. In this talk, we introduce NEC's SOL AI compiler, which seamlessly integrates with PyTorch, TensorFlow, ONNX, Numpy, and soon JAX. SOL provides a unified compiler engine for these frameworks, supporting both inference and training, while also enabling model export to standalone libraries with minimal dependencies. Designed for device-agnostic support and ease of maintenance, SOL requires no specific compiler support (e.g., OpenCL, SyCL, OpenMP, Triton, MLIR, ...) but can generate device tailored code with minimal coding effort. We will present SOL's key concepts and its device-agnostic design in this talk.
Biography: Nicolas Weber is a Principal Research Engineer at NEC Laboratories Europe, specializing in performance portability and automatic code optimization for diverse hardware platforms. He graduated in automated memory access optimizations for GPUs at TU Darmstadt. Since joining NEC in 2017, Nicolas has led and developed the SOL AI compiler. He's the main responsible for the PyTorch, TensorFlow, and ONNX integrations, as well as the NEC maintained hardware backends (X86 CPUs, NVIDIA GPUs, and NEC SX-Aurora). He has further been driving the development of the VEDA offloading API and other developments for the NEC SX-Aurora software eco-system. As a key accelerator programmer, he also advises NEC customers and business units on optimizing computational pipelines across various application domains.