Some comments on this that I will probably forget by tomorrow:
- They don't take into account bandwidth costs at all, either as a
strict upper bound or as a penalty function in their objective. One
could argue that their load capacity for each data center is a strict
upper bound on bandwidth, but it isn't complete (multiple links could
have different bandwidths coming in from different ISPs, etc)
- What is their annealing factor? Did they even tune this at all? Are
they really using the simulated annealing algorithm from 1983? (There
have been a few advances since then in the field of continuous
optimization...)
- How non-linear are the non-linear parts of their model? Can they be
approximated to a linear function? (If yes, the gains could outweigh a
simulated annealing approach)
- I like the separation of "Brown" vs. "Green" energy