The only way to really dig into this issue is wireshark/tcpdump on both ends to determine if calls arrive, what happens with them and what happens when they return to nginx (port closures, handshake error, timeout, etc.)
We serve loads of tomcat backends which are really slow sometimes and rarely see errors like these, yes they do occur but like once a week when too many users hit the backend (filling out their time-sheets).
Also have a good look at the settings of the network cards, energy saving mode, tcp offloading, qos, top/dis-responders, dns settings (like auto register this machine at....), unused netbios, calls to active directory which may interrupt, etc.
We serve loads of tomcat backends which are really slow sometimes and rarely see errors like these, yes they do occur but like once a week when too many users hit the backend (filling out their time-sheets).
Also have a good look at the settings of the network cards, energy saving mode, tcp offloading, qos, top/dis-responders, dns settings (like auto register this machine at....), unused netbios, calls to active directory which may interrupt, etc.
