nog.fi meeting 2025.06

Thanks to sponsors and volunteers, etc.

Daniel Wagner presents about most recent research activity at DE-CIX. Being part of the "Q-net-Q" consortium, latest advancements in the area of Quantum Key Distribution (QKD) in the context of data centers and at IXPs is being dicsussed. Having a slight touch of marketing style slides, the great cons and even potential show stoppers aren't dismissed; they are presented in a neutral and rational fashion to fuel an honest discussion.

By thinking of networking concepts—such as devices, cables, and configurations—as services, we’ve achieved higher levels of abstraction. Daily operations become faster and easier when performed at this abstract level. Services are carefully modeled, hiding complex details under the hood. This transforms each abstracted service into a ‘Lego block’—simple to understand and effortlessly connect with others. The result? A ‘design once, use repeatedly’ approach that boosts efficiency and scalability.

An analysis of what effects were visible using RIPE Atlas anchors in the region and elsewhere.

A perspective on how the power outage in Spain and Portugal on April 28 impacted Internet traffic, connection quality, and routing (announced IP address space) as seen through Cloudflare Radar.

The current RPKI framework using ROAs, only works to validate route origination.
To validate route propagation IP Transit providers will implement filters governing which routes can be propagated based on information from the IRRs (Internet Routing Registries) - in our region the RIPE DB.

This is a real life example of manipulation to the RIPE DB, which could have enabled a BGP Hijack to bypass transit provider filters and Route Origin Validation (ROV).

Was this a prelude to a BGP Hijack or a simple mistake? You be the judge.

In this presentation I will demonstrating about how we can do Cisco ACI configuration automatically by the help of Postman.
This presentation will be including practical demonstration that I will provide it based on the page 11 scenario in the following days.

Artificial Intelligence (AI) set new demands on current data center infrastructure, as it is undergoing a paradigm shift to be able to cope with the new demands possessed by AI clusters to support various use cases. AI clusters can be built in various sizes based on specific needs and workloads. This presentation will explore the data center network infrastructure evolution around the CLOS IP Fabric with the new requirements posed for building infrastructure for AI data center clusters and various types of workloads that these would host. The basic infrastructure needed for such are highspeed AI data center network, storage, and compute. Building Large Language Models (LLMs) requires centralized data, with demand for large bandwidth, while Inferencing is becoming more commonplace, which is driving a transition to a more distributed computing environment where traffic moves device-to-device and edge-to-core-to-cloud.

The new TREX Turku only supports RFC8950, and there's also a test route server for RFC8950 at TREX Tampere. This presentation discusses things to take into account when peering using RFC8950. There are also some examples.

This presentation investigates the proximity to a low Signal-to-Noise Ratio (SNR) threshold that can still maintain a tolerable Bit Error Rate (BER) in 100G / 400G / 800G network links. Additionally, we account for factors such as temperature and cable length to predict the duration for which a reliable network connection can be sustained between transceivers. The analysis, based on data retrieved using a Flexbox, focuses on comparing the reliability of coherent (16QAM) and non-coherent (PAM4) transceivers, with a detailed discussion on the implications of these technologies on network performance.

Generating 1 Tbps of traffic, especially if the packet size is small is challenging. Scaling it down to cheaper hardware is even trickier. In this talk I'll explain my path and all the challenges I had to build such load generator, from both hardware and software point of view. I'll talk about network card performance, problems that sub-NUMA clustering presents on modern hardware and explore their effects on the performance.

This presentation will give a practical showcase how peering related configurations can be automated using PeeringDB as a source of information. Elisa has been running automated peering configurations on their public IX sessions already for some time, so practical real-life examples will be given.

I made a tool that can take natural language request, forward it to an LLM. The LLM can provide network device commands for that request, given that LLM already knows the router model. Then that command is forwarded to the router, and lastly router feedback is forwarded to user. It is a bit underwhelming in real life, it is a cli tool written in python, and in very prototype-y phase. I want to present this idea to the community and want to know if this idea has any future.
All I have done till now will be available at https://github.com/jonathansaha/llm-router-interface.