I try to re-implement the algorithm Decima, published in SIGCOMM '19 (https://web.mit.edu/decima/). The official code its authors provide is written in tensorflow v1 and is a little confusing.
By the way, the authors do not open the codes for multi-resource scheduling (and the interface to Spark). I may implement it by myself in this repos.
It's also interesting to adapt Decima to K8S-based scheduling frameworks, such as Volcano (https://github.com/volcano-sh/volcano). However, we have to face many new challenges. For example, what can be scheduled is not the jobs themselves, but the pods which wrap them. Cross-server communication overhead also need to be taken into consideration.