I'm curious what backpocket optimizations you have from your past experiences which you pull out when a workload costs too much in the cloud. For me its.
* disable lower c-states
* enable adaptive coalesce in the network driver
* mount volumes with noatime
