with a minor comment on the code (extracting the cacheType variable to outer scope might not be required, though I only looked at the diff, I might be missing something in the bigger picture)
It got merged, and using our recently added nightly nugget feed I built a non-scientific experiment running this in a console app on my 4 core (HT) MacBook Pro. It squarely falls into micro optimisation bucket as any performance difference I noticed (ran a few scenarios, with varying number of iterations and proxy types) is almost negligible.