Abstract: Recent advances in computational methods for CSE and AI — such as computational drug discovery and LLMs — combined with new parallel computing hardware like GPUs, demand more than just parallel programming software (e.g., CUDA, MPI, Hadoop) to harness computing power for these applications. To ensure programmer productivity, parallel programming software relies on critical parallel computing system software — compilers, runtime systems, job schedulers — that enable performance, efficiency, and scalability, correctness, and resilience of the application. Moreover, as applications grow in complexity and computational demands evolve, applications require support of such system software across different parallel programming abstraction layers, from intra-node computation to cluster-level parallelism to cross-datacenter work orchestration. This keynote explores advances in system software frameworks that have enabled traditional scientific and emerging AI workloads. We will focus discussing systems software innovations across four layers of programming abstraction: (1) low-level parallel programming models (e.g. CUDA/HIP, OpenMP, MPI), (2) libraries for base-language-parallelism (e.g. Charm, Kokkos, C++ and Fortran StdPar), (3) Domain-Specific Libraries (DSLs) offering parallel programming (e.g. Triton, PyTorch, PetaBricks), and (4) AI agent and scientific workflow orchestration tools (e.g. NVIDIA’s AgentIQ, Pegasus). In doing so, we will share open-source parallel computing systems software developed from our computer science and HPC research that has supported production workloads in science and AI at national labs and AI companies, and we will discuss how such systems software used in these organizations has impacted systems software for today’s rapidly evolving computing needs to enable high-demand genAI in industry. We will close with a perspective of how current parallel computing system software will extend to emerging and future application domains such as robotics, blockchain, and other domains involving real-time computing systems.

Biography: Vivek Kale is a Principal Member of Technical Staff at AMD, working on profiling and debugging tools and AI-assisted profiling tools for the ROCm GPU software stack. Prior to this, he was at Sandia National Laboratories—California, where he led R&D of parallel computing systems software for science, engineering, and AI workloads vital to the United States Department of Energy, with an emphasis on profiling+debugging tools for GPU programming. He earned his PhD in Computer Science from the University of Illinois at Urbana-Champaign in 2015. Vivek’s work revolves around driving software development of tooling library infrastructure and capabilities, ensuring the tooling library’s interoperability with MPI and cluster-level job schedulers, and influencing enhancements to GPU programming models such as Kokkos, OpenMP and HIP to facilitate tooling for customer application use cases. Vivek's work has also concentrated on AI-assisted tools, particularly leveraging LLMs for source-to-source GPU program translation tools and developing ML-guided auto-tuning tools for Kokkos applications.
https://vlkale.github.io/

Abstract: The recent revolution in LLMs has brought us closer to new and challenging dimensions of AI. How has this been achieved by leveraging the cloud and big data? A critical and challenging analysis is required here. Furthermore, the transition from the cloud to the edge has given rise to distributed intelligence, including federated learning. The current trend towards agents, new workflows, agentic AI, and more specialised LLM AI models is speeding up the process of an effective Distributed Intelligence. Let's take a closer look.

Biography: Massimo Villari is Full Professor in Computer Science at University of Messina (Italy). He is actively working as IT Security and Distributed Systems Analyst in Cloud and Edge Computing, virtualization and Storage, Federated Learning and one of the creators of Osmotic Computing Paradigm. For the EU Projects “RESERVOIR” he leaded the IT security activities of the whole project. For the EU Project “VISION-CLOUD” and “H2020-BEACON”, he covered the role of architectural designer for UniME. He was Scientific ICT Responsible in the EU Project frontierCities, the Accelerator of FIWARE on Smart Cities – Smart Mobility. He is strongly involved in EU Innovation initiatives, and also, he covers the role of EU external expert. Currently, he is Scientific ICT Responsible for UniME of “TEMA” and “NEUROKIT-2E” Horizon Europe Projects. He is co-author of more of 390 scientific publications and patents in Cloud Computing (Cloud Federation), Distributed Systems, Wireless Network, Network Security, Cloud Security and Cloud, Edge and IoTs, and recently in Osmotic Computing and AI, Federeated Learning and Edge AI. He was General Chair of ESOCC 2015 and IEEE-ISCC 2016. In 2014 he was recognized by an independent assessment (IEEE Cloud Computing Transaction, Issue April 2014) as one of World-Wide active scientific researchers, top 27 classification, in Cloud Computing Area. He was General Chair of IEEE-ICFEC 2019 and General Co-Chair in IEEE-CCGRID2022. He was also General Chair of IEEE ISCC2022 and IEEE ISCC2024. Since 2018 he is member of SC of IEEE-ISCC. He is General Co-Chair in IEEE-CCGRID2026 in Sydney. He is in the Stanford World’s Top 2% Scientists list of the Stanford University in Computer Science. He is a Co-Founder of UniME Spin-Off Alma Digit S.R.L since 2017. He was Rector delegate on ICT for whole University of Messina and Academic and Consultant for the Messina Municipality in the Area of Smart Cities. Currently he is Head of fcrlab group @unime, and External European Expert on AI, Robotics, Computing Continuum, IoT and Cyber Security @EU, and the UK Edge AI Hub Advisory Board Member and the Head of Data Science Schools.
https://archivio.unime.it/it/persona/massimo-villari/biografema

Abstract:

Biography:
https://istc-arsa.iisp.gatech.edu/pages/michael-kounavis.html

Abstract: Large language models have taken the world by storm. They are undoubtedly effective and have been able to perform tasks seemingly impossible for automated system only a few years ago. However, there is a limit to their capabilities and effectiveness. In this talk I provide an overview of what could perhaps be construed as limitations of these models. I provide my opinion on what would most likely be the approach that would be taken as we move forward in this domain.

Biography: Abhishek Srivastava is a Professor of Computer Science and Engineering at IIT Indore. His research spans varied domains but most specifically is placed at the intersection of physical systems and artificial intelligence. His group is dwelling into the possibilities of extending the realms of AI and machine learning to the physical world, an endeavour often labelled as physical AI.

Abhishek has been with IIT Indore since 2012 and has served at various administrative positions including Dean of Faculty Affairs, Dean of Student Affairs, and Head of Computer Science and Engineering. He was an Assistant Professor at the Rose-Hulman Institute of Technology, Terre Haute, USA; and has a Ph.D. in Computing Science from the University of Alberta, Canada.