if your nodes are going to share a physical CPU, then I think it's probably best to just have one pod per CPU, otherwise you could end up with some really bad scheduler contention and priority inversion