Hajed Khlifi
Principal Solutions Architect
Luxembourg
Actions
I architect and deploy production ready, secure, large-scale HPC and AI platforms on air-gapped, sovereign, hybrid and public clouds using Kubernetes, OpenShift, Slurm, NVIDIA HGX and DGX, AWS, GCP and full enterprise tooling. I build GPU clusters, HPC schedulers and LLM inference platforms for regulated enterprise markets, from banking, fintech, Government to life sciences and reasearch (I am super flexibale).
Area of Expertise
Topics
Slurm Already Solved This: Teaching Kueue to Schedule License Tokens on Kubernetes
In biopharma research a job usually needs more than a GPU. Docking and free-energy tools check out license tokens from a server when they start and Kubernetes doesn't know that server exists. The scheduler starts the job, the checkout fails, and an H100 sits idle and billed while the solver retries. The queue is full, utilization is low and no dashboard explains why.
Slurm handled this years ago by treating licenses as one shared pool and making jobs wait until tokens are free. This talk brings the same idea to Kubernetes. Attendees will see how queueing works in Kueue / Multi-Kueeue from a suspended Job to running pods and how a Prometheus exporter, a small quota reconciler operator and a custom AdmissionCheck make jobs wait for licenses instead of failing.
It is for platform and HPC engineers running licensed software on Kubernetes, and ends with the mistakes that break this setup and the alerts that catch them.
GPU is not Monolithic : Packing LLMs with MIGs on Kubernetes
Most of the LLM workloads now are deployed on Kubernetes clusters with GPU nodes and let's be honest this is the most expensive resource in the cluster. Currently, using GPUs in passthrough mode locks a single model to an entire GPU, leading to severe underutilization (~30%). In this talk I will explain how to manage GPU resources in an efficient way and attendees will understand how GPU cards are configured in a Kubernetes cluster, what is the difference between the three main Nvidia GPU installation modes: Passthrough, vGPU and MIG and how everything works behind the scene. I will demonstrate how Multi instance GPUs are the best solution for packing LLMs on Kubernetes and how it should be used in an advanced case scenario like packing multiple LLMs in the same cluster sharing the same GPUs without causing the noisy neighbor problem.
Breaking Barriers with Dapr: Simplified and Portable Cloud-Native Architectures
Microservices are often built with different technologies, leading to complexity in management and integration. This session demonstrates how Dapr simplifies multi-stack microservices across cloud environments. We will explore an e-commerce application built using Golang, Java, Python, and Vue.js, that we are going to deploy live (1) on premises (Docker-Compose), then on (2) AWS, then (3) Azure and finally orchestrated in a (4) multi-cloud setup. Attendees will see how Dapr’s features (service invocation, state management, pub/sub) abstract complexities, enable easy development and migration across environments. This session highlights how Dapr’s "Lift and Shift" approach facilitates seamless cloud transitions without re-architecture, making it an ideal solution for modern, multi-stack microservices.
WeAreDevelopers World Congress 2026 - Europe Sessionize Event
Cloud Native Days Italy 2026 Sessionize Event
Container Days London Sessionize Event
ContainerDays Conference 2025 Sessionize Event
KCD New York 2025 Sessionize Event
Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.
Jump to top