Old habits and myths die hard. Conventional wisdom asserts that UDP is better because it has lower overhead; then conventional wisdom suggests that you tune the buffer sizes to improve performance. On the face of things that would seem to work but once the the write size exceeds the max packet size, NFS delivers the packet by using multiple packets. Sending multiple packets triggers the issue because dropping just one UDP packet means the whole buffer must be resent. Contrast with TCP: yes the packet header is larger so less data can be sent and yes the receiving side has to ack each packet. But: with TCP if a packet gets dropped, only that packet needs to be resent; with a modern TCP stack the kernel will constantly adjust the window size to make the best use the available bandwidth. In other words NFS over tcp will automatically tune the buffer sizes for the current conditions.
Yet another open source journal