The traditional LinuxDays 2025 conference took place in Prague on the weekend of October 4–5, attracting experts and enthusiasts in open-source software, Linux, and related technologies. e-INFRA CZ was once again among the exhibitors, with our specialists answering visitors’ questions about high-performance computing (HPC), data storage and management, network security, and artificial intelligence tools, including large language models (LLMs) operated within the national e-INFRA CZ infrastructure.
The program also featured two expert talks presented by e-INFRA CZ representatives:
Large Language Models in Practice
Lukáš Hejtmánek from CERIT-SC delivered a presentation titled “Large Language Models in Practice: Experiences from e-INFRA CZ Operations.”
He shared more than six months of hands-on experience running in-house instances of large language models designed for scientific research. He explained why e-INFRA CZ chose to operate its own LLMs rather than rely on commercial solutions, and discussed theoretical aspects such as the relationship between parameter count and model performance, the principles of quantization, and the role of KV cache.
The main part of the presentation focused on practical experiences, optimizations, and challenges encountered by the team—such as memory allocation issues in certain Linux kernel versions or running multiple models on a single GPU. The talk demonstrated that large models can be efficiently operated even on moderately powerful hardware.
XDP – A Programmable Network Stack
The second talk was given by Richard Hyroš from CESNET on the topic “XDP – A Programmable Network Stack.”
He introduced XDP (eXpress Data Path), a technology leveraging eBPF (extended Berkeley Packet Filter) integrated into the Linux kernel to significantly boost network stack throughput. XDP enables packet processing directly within the kernel before packets reach higher layers, dramatically reducing latency and increasing performance. Hyroš also discussed advanced modes such as AF_XDP (zero copy), which allow bypassing the traditional network stack and efficiently directing packets straight to user-space applications—an alternative to frameworks like DPDK.
