I eventually managed to configure TeamCity to run a very small set of
(non-global) inspections on our code base. This runs in about 20-30
minutes in 1G memory on a 32-bit machine with build 1784. However, the
same set of inspections takes up to 4 hours with the same configuration
using Agra (3654)!
It looks like Agra uses a lot more memory, but I'm not sure this is just
3 hours of garbage collection. I have CPU and telemetry snapshots from
both builds, are you interested?