This started off as a comment but got a bit long.
There is no such thing as connection pooling in HTTP. Transaction pipelining is possible (see HTTP keepalives) but this has nothing to do with the process engine. (SPDY also allows for transaction multiplexing - but that's something different again). Adding servers (appropriately configured) will decrease connection re-use hence worsening your performance.
In addition to what appears to be a flawed premise and incorrect assumptions, you've omitted a lot of information which is relevant - What OS is running at both ends? Are you using SSL? What is your SSL session timeout? Are you using a shared SSL session cache?
If network latency is the big problem here, then running the HTTP through a VPN or IPSEC instead of SSL will improve the situation (no need for SSL [re-]negotiation for every request but you still have a TCP handshake).
If SSL is not a consideration, then make sure you've got window scaling set up properly with appropriate RWINs at both ends. Most OS will cache the slow start threshold across connections which means that it shouldn't have too much impact unless connections between the end-points are relatively infrequent - in which case you might want to set a higher default.