Well, I thought it might be GC related also. It behaves similarly to how you would expect a GC pause to behave (several fast responses, then one that takes 5 seconds to come back).
I think that would be plausible if parts of the managed heap were swapped out to spinning-rust HDDs. Fortunately, RAM is plentiful in our case.
