When tcp_tw_reuse is enabled the kernel can decide to use the sockets in TIME_WAIT, before they expire or they are closed by the clients.
This is a problem though, because the connection could still be used by the client and therefore there could be some collisions regarding the TCP sequence numbers, specially on high traffic servers. The kernel can try to avoid this collision with a technique called PAWS (protection against wrapped sequence numbers: rfc1323). Unfortunately PAWS works only with tcp_timestamps enabled on both sides (client and server). tcp_timestamps has also an overhead and therefore it is normally disabled on servers with a high traffic, leading to potential problems.
About tcp_tw_recycle, when it is enabled, it forces the verification of this tcp_timestamp. So in case of NAT, multiple clients will send different tcp timestamp to the server, to the same mapped connection which points to the TIME_WAIT socket, and because the tcp timestamp are different then the packets will be dropped by the kernel. This is the reason why it is not a good thing to enable tcp_tw_recycle when you use a load balancer or in case of NAT.
A good practice is to enable tcp_tw_reuse (instead of tcp_tw_recycle), to make sure tcp_timestamp is enabled and to decrease the size of the tcp timestamp with tcp_timewait_len.