Any chance you have pointers for public trace files? Its interesting to see the behaviors under different workload patterns.
Adding a cost model (GDSF, et al) is too specialized that I haven't spent time looking in that domain. From your hints that sounds like the unexplored area that your work successfully optimized.