BEAM has some heuristics where it can switch to a busy wait in some cases in an attempt to reduce latency. This is tunable with the +sbwt argument to erl; but the default means it's hard to use the system CPU % as a load indicator.

It would be interesting to hear more about this because with CPU spiking like that I would have expected more variability in the response time...but quite the opposite it was the most consistent which is one of the big things that the BEAM aims for.

I wonder if somebody who's more of an expert on the BEAM could chime into why that might be. More activity in the scheduler maybe?