A technique for recovering from a socket error at a computer host. The host opens two sockets for communicating with respective sockets at another host. Failure recovery processes are provided for a client host on its server-initiated receive socket and its client-initiated send socket, and for a server host on its server-initiated receive socket and its client-initiated send socket. A failure may be cause by an operating system error, a lack of communication at the socket, or removal or failure of a communication medium such as a cable. When a failure condition is detected at one socket, the host attempts to use the other socket. If the other socket can be used, an attempt is made to reconnect the failed socket. An internal state is set at the host denoting a normal mode or an attempt to recover mode.