Build Logs & Dev Diaries

Much things got clarified. But I still don't know how we will orchestrate the LLM and move its cache through the 5G network. I still genuinely don't think there will be any benefit to using UPF for the control/management plane. Need to dig into more 3GPP standards and more papers to be argued next week.

But I'm kinda relieved that all that hassle is actually part of the project. For my thesis, all I need to do is just find the most suitable node. Moving all those through the 5G network is part of the big project. But I'm still worried because only two nodes are capable of running it. I'm afraid there won't be much difference from the default Kubernetes.

Containerized vLLM with Qwen model. Took up most of my time today. Went home at almost midnight because of many issues. First, the sizes were too big, and I had to resize the LVM for both machines. Then, the NVIDIA GPU was not detected inside the pods, and I had to install some toolkit beforehand. The Docker build time was another frustrating matter. But all of that is worth it. All safe and sound. Hope everything works out all right tomorrow.

I truly don't know what I'm actually doing in my thesis. I'm fucked. No one told me that the research phase is the hardest part. I think I might questioning my whole existence because of it.

Hi! It's been a while. I caught up on a lot of things (it's just my thesis, honestly). The scope is, I believe, much narrower and clearer by now. Yesterday was supposed to be the fixation of the evaluation method that will be used in the research. But since the time is so limited, only one of us was able to present it, and the rest can just send the pptx file to be reviewed by the lecturers. Since it hasn't been reviewed yet, I don't know what else to do.

For the work itself, resource telemetry from the node is already done. Next, I might try to rebuild the core to be able to get the UE data, to determine where that particular request comes from. After that, the next step would be to put the algorithm into code. Might be the hardest part of all. But if it's done, all that's left is just integrating it with Kubernetes and testing it.

But truthfully speaking, testing is the thing I am most afraid of. The implementation code itself, I believe, is just an easy part (I hope so). But knowing whether what I test is valid or not is just another thing.

Revised the 1st chapter. I believe it's more coherent than it was. For the testbed, I still don't know when I can start implementing all those fancy telemetry tools.

Went to campus this morning to assist with a practicum. It was a cybersecurity class. I myself don't really have much interest in it, but it shouldn't hurt to try, right? After that short class, went to the lab to do small talk with others, then went home and slept. Not sure what to do afterwards, then I decided to review my thesis proposal that is due to be reviewed tomorrow morning. I hope that not much will change. I'm already sick of it.

Got summoned to the Lab for a meeting regarding lecture modules. It was then cancelled after a miscommunication. Ended up reformatting the ICampN module. It was a pain in the ass to reformat all those pages. I don't get why it is so hard for people to follow the template we've been given in the first place.

The meeting went fine. Probably gonna need to adjust my placement approach. Need to consider using a cleaner approach like Kustomize, though yes, it's more static than my approach.

Found a similar approach used by SoftBank. They use AITRAS framework. I thought that they were already working on my research. Glad they solve different things, so I can still stick to what I'm doing. They employed dynamic scoring for managing multi cluster. They delegate what services to be deployed, not granularly on the node level, but just on the cluster level, and let the scheduler handle the rest.