It's surprisingly insightful to monitor system load by the UPS load. The setup on the left needs some tuning.
@pixel Facebook did the reverse, using current measurements to train a model to predict power draw based on system load. The goal was to be able to avoid having to overprovision for worst-case power draw by throttling servers before they blew breakers.
(I originally thought I'd be violating my NDA by telling you that much, but it turns out they've published a paper about it: https://www.researchgate.net/publication/309142298_Dynamo_Facebook's_Data_Center-Wide_Power_Management_System)
Cool to know. By now the setups are equal, as most of the data got cached into memory and there is less disk io. Don't tell my boss that he can judge my productivity by the electric bill!
A beta setup of a Mastodon instance primary for family and friends.