Comment 12 for bug 361733

A while back, I changed my doctoral specialization to computer architecture. But prior to that, I was studying AI, and as I think about this, this problem screams for a reinforcement learning approach. For N drives, there are 2N+1 actions, those being read to each drive, write to each drive, and possibly do-nothing (even if there is work to do). If we can map past actions, knowledge of the consequences of those actions, and information about time since last action to some reasonable number of "environment states," then it would be straightforward to implement a completely adaptive online optimization algorithm. It could also be tunable to favor throughput, latency, and even energy to different degrees. The general idea is to use measurements of actual performance to judge the effectiveness of past actions in achieving those goals and integrate that into future decisions. If we did it right, the algorithm would automatically adapt to differences in drive performance characteristics. Mind you, this might only get incremental improvements in performance over the more naive approaches. But I recall reading one paper about a DRAM controller scheduler that achieved a 22% throughput improvement over the best static method available at the time. (And as I continue to think about this, perhaps the best place to implement this first would be for reordering accesses to single drives.)

Sorry about the rant. :)