4 days of lectures and hands-on sessions

9h00 Plenary Session

Building the High-Performance Core of AI Factories

Farouk Mansouri, LuxProvide

AI factories must be designed to handle the most demanding stages of the machine learning lifecycle—data ingestion, preprocessing, and large-scale model training. These early phases place extreme demands on infrastructure, from GPU-accelerated supercomputers and high-throughput storage to low-latency interconnects and distributed data pipelines. This session explores how to architect and optimize dynamic, high-performance environments capable of processing massive datasets, orchestrating parallel training jobs, and scaling resources efficiently. Participants will gain practical insights into overcoming data bottlenecks, maximizing hardware utilization, and integrating HPC and cloud resources to deliver speed, scalability, and cost-efficiency at the start of the AI production chain.

9h45 Plenary Session

- Evolving Cloud Infrastructures for AI Workloads - High-Performance Foundations for AI Training

Clara Ulken, GCore

At GCore, Clara is leading cross-functional initiatives that align engineering, product, and business strategies to deliver scalable, high-impact technology solutions.

AI development begins with workloads that demand extreme performance. Training large language models and generative AI systems requires GPU acceleration, low-latency interconnects, and scalable storage to handle massive datasets. This session explores how Gcore’s GPU Cloud—powered by NVIDIA A100, H100, and H200 instances with InfiniBand networking—delivers HPC-grade capabilities in a flexible, cloud-native environment. We will look at distributed training strategies, mixed precision techniques, and orchestration tools that transform supercloud-class resources into elastic infrastructures for enterprise AI.

10h30

** 30 minute break **

11h00 Plenary Session

Delivering Agility: From Trained Models to Scalable AI Services

Farouk Mansouri, LuxProvide

Once AI models are trained, the challenge shifts to delivering them as fast, reliable, and cost-effective services. Inference workloads require entirely different priorities—low latency, high availability, and seamless integration into production systems—often through containerized microservices, auto-scaling cloud platforms, or edge computing. This session examines how to bridge the performance–agility gap, from compressing and optimizing models for deployment to building resilient MLOps pipelines for monitoring, retraining, and governance. Real-world patterns and case studies will illustrate how to move from HPC-heavy training to lightweight, scalable inference while maintaining cost control, compliance, and operational excellence.

11h45 Plenary Session

Evolving Cloud Infrastructures for AI Workloads – Agility at Scale: Deploying AI Everywhere

Clara Ulken, GCore

AI is no longer confined to research or centralized datacenters, but it increasingly powers real-time, interactive experiences where every millisecond counts. This session focuses on Gcore’s Everywhere Inference platform, which brings models closer to end users, enabling ultra-low latency, reducing unnecessary backhaul traffic, and supporting regional data handling requirements. Gcore has extended its global edge infrastructure with GPU-powered Points of Presence worldwide, creating a platform designed for workloads where speed and locality make a measurable difference. With Kubernetes integration, autoscaling, and support for hybrid deployments, enterprises can deploy optimized inference services that adapt in real time to diverse workloads. From finance to conversational AI, fraud detection and live content personalization, to immersive gaming and AR/VR, inference at the edge delivers tangible improvements in performance and user experience.

13h00

** lunch break **

14h30 Hands-on Session

Hands-on Session. Participant will have the opportunity to access a remote super-computer to run tests and experiments

Part 1 – HPC Training/Finetuning of an LLM

Goal: Show participants how to run and scale a training job on an HPC system. (45 minutes)

1. Introduction to HPC Infrastructure

o Overview of the HPC cluster architecture (login nodes, compute nodes, GPU nodes) o SLURM basics: job submission, partitions, scheduling policies o Storage layout and data transfer tips

2. Exploration of Performance Tools

o htop, nvidia-smi, ibstat for InfiniBand, iostat for disk I/O o Brief on nvtop or similar GPU monitoring tools o How to interpret load, memory, and network usage

3. Simple Run

o Launch a minimal LLM training job (small dataset, reduced parameters) o Show SLURM job submission (sbatch), log checking (squeue, sacct) o Walk through model directory structure and outputs

4. Run at Scale

o Increase dataset/model size and GPU count o Demonstrate distributed training (PyTorch DDP, DeepSpeed, or Megatron-LM) o Monitor scaling behavior and GPU utilization in real-time

5. Monitoring & Optimization

o Detect bottlenecks (GPU idle time, I/O wait, network congestion) o Adjust batch size, precision (FP32 vs FP16), and parallelism strategies o Quick discussion: trade-offs between speed and accuracy

14h45

** 10 minute break **

Part 2 – Kubernetes Deployment of the Trained Model

Goal: Deploy models as a scalable inference service (45 minutes)

1. Introduction to Kubernetes Infrastructure

o K8s architecture: master, worker nodes, pods, services, ingress o Overview of deployment options: cloud K8s, on-prem, hybrid o Brief on container images and registries

2. Exploration of Performance Tools

o kubectl top nodes/pods for resource monitoring o Logs (kubectl logs), kubectl describe for debugging o K8s dashboard or Lens for visual inspection

3. Simple Run

o Deploy model as a single replica pod with REST API (FastAPI/Flask) o Expose via NodePort or port-forwarding o Test with a sample query

4. Run at Scale

o Scale replicas using kubectl scale or HPA (Horizontal Pod Autoscaler) o Demonstrate load testing (e.g., hey or ab command) o Show autoscaling behavior under increasing load

5. Monitoring & Optimization

o Identify bottlenecks: CPU/GPU constraints, network latency o Optimize container startup, model loading time, batch inference o Discuss cost/performance trade-offs in scaling

15h40

** 10 minute break **

Part 3 – End-to-End Challenge & Wrap-Up

Goal: Apply both HPC and K8s learnings to simulate a full AI factory pipeline (45 minutes)

1. End-to-End Integration

o Take the trained model from HPC output o Package it into a Docker image o Push image to registry for K8s deployment

2. Performance Challenge

o Teams run the model at scale and optimize throughput & latency o Compare scaling behavior between HPC training and K8s inference o Introduce optional constraints (cost cap, latency target)

3. Wrap-Up & Key Takeaways

o Recap performance/agility lessons from lecture o Share best practices cheat-sheet for HPC and K8s operations o Q&+A + feedback

9h00 Plenary Session

Containers/Docker

Giorgos Saloustros, FORTH

Containers are now the standard for application deployment. By packaging an application's code and all its dependencies into a single, isolated unit, they ensure consistent behavior across any environment. This makes applications portable and much easier to develop and share. This session will introduce the technology and present the reason of its ubiquitous success as well as the challenges remaining to be addressed

9h45 Plenary Session

Orchestration/Kubernetes

Antony Chazapis

Orchestration is essential for managing containers at scale. Without it, manually deploying and scaling applications quickly becomes unmanageable. Kubernetes automates these tasks, handling everything from load balancing to self-healing, so your applications are always available. Orchestration is the fundamental concept to build a service architecture out of a set of individual containers

10h30

** 30 minute Coffee break **

11h00 Plenary Session

Worflows

Antony Chazapis

A workflow is a sequence of computational steps. It structures complex processes where the output of one step becomes the input for the next. This is fundamental for data processing and machine learning, making these pipelines repeatable, transparent, and scalable. This session will present the key concept the main pitfalls and the most relevant technologies to implement a workflow.

11h45 Plenary Session

Argo

Lefteris Vasilakis and Antonis Tapanlis

Argo is a suite of tools for building and managing workflows directly on Kubernetes. Argo Workflows allows you to define complex, multi-step pipelines natively within your cluster. It's an ideal technology for running large-scale data and machine learning tasks. This session is the concluing stage of the morning session where after presentation of the individual containers, the need to organize and share resources and orchestration, a fully operating workflow can be implemented and deployed.

• Slides of the handson Argo session

13h00

** lunch break **

14h30 Hands-on Session

Hands-on Session. Participant will have the opportunity to access a remote super-computer to run tests and experiments

Knot

Giorgos Saloustros/Antony Chazapis

prerequisite:

access to an HPC Cluster will be provided by DaFab

Outline of the sessions

The hands-on session will start with Docker containerization for packaging and managing applications in reproducible environments, followed by Kubernetes fundamentals for orchestrating and scaling workloads. The session will also introduce Knot a novel open-source orchestrator developed by FORTH harnessing both Kubernetes workflow and Slurm the HPC resource manager.

16h00

** 30 minute Coffee break **

16h30 Hands-on Session

Hands-on Session. Participant will have the opportunity to access a remote super-computer to run tests and experiments

hands-on create your own workflow with Knot/Argo

Lefteris Vasilakis/Antonis Tapanlis

prerequisite:

access to an HPC Cluster will be provided by DaFab

Outline of the sessions

Participants will explore further workflow design principles, with a deep dive into Argo Workflows for managing complex, automated pipelines and the way Argo and knot can interoperate to provide seamless workflow deployment on different infrastructures, either cloud based on HPC.

• Source code of the Argo workflows

9h00 Plenary Session

The AI Factory for Earth

Stephanie Giard, DDN

Stephanie is a Senior Product Marketing Manager at DDN. She brings extensive experience in Earth Observation from her roles at Planet, Airbus Defence & Space, Boeing, and Ubotica. Based in Fort Collins, CO, she holds a Master’s degree in Remote Sensing and GIS from Boston University.

There is an important on-going evolution of the Earth Observation data market. Originally the data market was driven by data acquisition, data acquisition cost and complexity. The Earth Observation community has been extremely successful in its effort to democratize data acquisition, as a result the community is drowning in data and starving for insight. The bottleneck is not satellites or sensors—it's data infrastructure. In this talk we will discuss this on-going evolution and the key data infrastructure technologies required to address this problem specifically in the time of AI.

9h45 Plenary Session

Data and Performance: Key factors and Parameters

Jean-Thomas Acquaviva, DDN

In this presentation we will present the traditional pitfall in data logistic, from management to exploitation. We will introduce some useful metric to assess the scalability and the relevance of a data infrastructure

• Slides of Performance presentation

10h30

** 30 minute break **

11h00 Plenary Session

Metadata Catalog: The needle and the haystack

Dimitrios Xenakis, CERN

Earth Observation today produces petabyte‑scale streams scattered across multiple mirrors and portals. The result: finding the right scene is often harder than processing it. This session introduces a practical path from needle to hits. We’ll start by mapping today’s discovery landscape and give a concise primer on STAC, identifying its strengths but also where it falls short. Then we’ll show how the DaFab AI project uses CERN’s Rucio to unify source and AI‑derived metadata, and how the Rucio Enhanced Filter (REF) is designed to improve the metadata discovery.

11h45 Plenary Session

Observability and data performance at the time of AI

Jean-Thomas Acquaviva, DDN

In this presentation we will review suitable tools and methodology to observe I/O performance and tune workload to suppress potential data bottleneck. This session is an introduction of the hands-on tutorial in the afternoon

• Slides of Observability presentation

10h30

** 30 minute break **

11h00 Plenary Session

Metadata Catalog: The needle and the haystack

Dimitrios Xenakis, CERN

Earth Observation today produces petabyte‑scale streams scattered across multiple mirrors and portals. The result: finding the right scene is often harder than processing it. This session introduces a practical path from needle to hits. We’ll start by mapping today’s discovery landscape and give a concise primer on STAC, identifying its strengths but also where it falls short. Then we’ll show how the DaFab AI project uses CERN’s Rucio to unify source and AI‑derived metadata, and how the Rucio Enhanced Filter (REF) is designed to improve the metadata discovery.

11h45 Plenary Session

13h00

** lunch break **

14h30 Hands-on Session

Hands-on Session. Participant will have the opportunity to access a remote super-computer to run tests and experiments

prerequisite:

Installation of the VPN software Wireguard on a laptop. Availability of ssh

Git and Docker available on the laptop.

Outline of the sessions

Performance measurements with Darshan. Darshan is data profiling https://github.com/darshan-hpc/darshan tool used to trace applications behavior and report I/O patterns. In this session we will analyse mysterious applications to discover potential I/O bottlenecks using Darshan tracing tools.