Just shooting in the dark here, but:

Can you get a complete packet trace of packets to and from the
crashing host?  Either with a dedicated capture machine+hub, or using
a managed switch and "port mirroring" or whatever your switch vendor
calls it?

Seeing the TCP dynamics in conjuction with the stack traceback would
shed a lot more insight on what's going wrong.  OTOH, its a quite lot
to ask...