Project Lightning Talk: How To Run Kubernetes Pods On My Slurm-Based HPC Center - Diego Ciangottini

This talk introduces Interlink, a tool designed to seamlessly connect Kubernetes with high-performance computing (HPC) resources managed by systems like Slurm. The speaker explains that Interlink functions as a translator, enabling applications running on Kubernetes to effectively utilize supercomputers without requiring extensive administrative privileges or custom resource definitions. He describes the core components of Interlink, including a virtual kubelet provider, an Interlink server on the HPC side, and a network plugin. The session provides a demo showcasing how users can define their notebook requirements, target specific resources, and create Kubernetes pods that leverage the computational power of supercomputers such as Leonardo in Italy. The speaker invites the audience to learn more and see the tool in action at the kiosk on Wednesday.

Yeah, I'm a researcher at the uh National Institute of Nuclear Physics in Italy. But today I'm going to introduce Interlink uh as one of the maintainers. So what's this tool for? First of all, uh well, you can imagine that you create a pod on Kubernetes and with it manifest either you or your framework and you want to translate this into something that can run on a supercomput

and the supercomputer might be managed in a different way like with a Zur uh batch system and you want something a translator in the middle. Um this is the role of interlink uh and the ambition of interlink to make this as seamless as possible. What I mean by seamless well no CRDs plain pods definition and no root administrative network privileges to do all the the magic around.

So there are two disclaimer though the first one is that we are not talking about running slurm under kubernetes. So man managing u demons through kubernetes and we are not either talking about a situation where you have a bunch of nodes and you want to manage slurm and kubernetes side by side. Okay, we are really talking about something that is remotely access slurm based resources from your

own kubernetes that can run on your laptop on your cloud or whatever. So it can be closely remote meaning in the same center but with some network segregation or distant like right very distant from a plane running K3S. Hope you like this picture because it cost me to to make it works on the plane with the Wi-Fi. So yeah, wasn't a good idea. Um the requirements first

of all yeah um decently recent Kubernetes uh version and if you want the full-fledged capabilities of the system you might want also to have an ingress with wild cards or whatever is the equivalent approach on Zoom side on the nodes you have to run a container runtime and share the file system across nodes and at the edge node where you have to install one of the components

the network capability are caveat that uh should use unshare command on the nodes. So interlink it's composed by three main core components. The first one is a cublet virtual cublet provider for interlinker installed on the kubernetes cluster and the other one is an interlink server on the remote HPC side can be a login node or a dedicated machine on the edge of your supercomputer and then you

need a network plug plugin if you want your pods offloaded to the HPC to be able to talk with the internal pod network of your kubernetes cluster and that's all then you need just to target the virtual node created in this way in your pod just like an any normal pod and you're done. Today's demo will be very brief. We will have more on the kiosk as

I tell you. Uh we run on a supercomput also in Italy that is called Leonardo. So was done in collabor collaboration with them and we are basically running a cubeflow notebook can be a pipeline can be whatever everywhere uh they are pod in the end and let's let's see it briefly so you start and you define your uh notebook requirements and one cool thing I want you

to focus is that then in the configuration you can target the proper nodes and you uh can have multiple nodes, multiple supercomputers if you will. And once targeted, you sh you launch and it creates pods, right? Normal pods. On this side though, there is the login node of the Zoom HPC center and you can see that a job has been created. So you go ahead, you have

a full integration. You see here in a moment sorry to okay in a moment you can see here you have access to the logs from web interface and when you are done you have your notebook with GPUs of the supercomputer available. This is uh just an example. If you want to know more we can have liveing and you can see it in practice also on your laptop

how it works. You can join us at the kiosk on Wednesday from 10 to 1:30. Thank you.