I concur with the discussion so far; once you are into the multi-tor-daemon deployments the "tuning" becomes rather organic.

I would look at network throughput on the tor nodes which are serving as reverse-proxies, and correlate that against load.

Frankly: Facebook currently delivers its entire onion service through 2 daemons per each of 3 onion addresses (www, cdn, sbx) for six daemons total; the chokepoint is far more likely to be the content-delivery backend.

-a