When testing the performance CouchDB it's apparent that the main hurdle here, is the HTTP create + tear down. Hence, anything that can re-use the HTTP connections will boost performance.
I did some tests on what it meant for performance that I changed various things in my HTTP client setup and the findings were all not that cheerful - albeit interesting. no real difference between using multi threaded HttpConnectionManager and the standard one (when running against one host) as HTTPClient will still re-use the client's connection if the consecutive request goes to the same host (yes, even if you've called releaseConnection).
I gained at least one second (down from 8.5 to 7-8 seconds) by
commons-io/IOUtils to consume it.
There was no real gain to use many threads fetching statistical data in concurrent threads as the main bottle neck was the initial GET of statistical data.
CouchDB doesn't support gzip encoding of the content. When querying large datasets, this becomes a problem.
Putting nginx in front which did gzipping of the content. This turned out not to help significantly, running pure curl from the command line only showed a mere ~2 second improvement from 7 to over 5 second. The total running time including Java space (single connector).
In Java space, it was a nice discovery that HTTP client 4.x has built in support for gzip (if you use a certain implementation of HttpClient), the sad thing about it is that the API has changed quite a bit from 3 to 4 and it'll take you a mouthfull to change your code to use it.