1. You need to recognize the opportunities exist in the first place.
2. You need a global controller that can aggregate and optimize for a global solution (and the global solution might not necessarily simply to maximize the aggregate throughput, but there might be other factors into account), which may involve some algorithmic design (in some cases, you need to design new algorithms).
3. You need to justify that global controller gives you a superior solution compared to locally greedy solution. As in this article, a global solution gives you about 3% improvement compared to the local controller, and the local controller algorithm is substantially easier to write.
Background: in my previous job at Meta, I wrote such a global control algorithm for controlling the rate of data going in and out each data center. It involved some really interesting algorithmic design.