18.10. Using NFS over TCP
The default transport protocol for NFSv4 is TCP; however, the Red Hat Enterprise Linux 5 kernel includes support for NFS over UDP. To use NFS over UDP, include the -o udp option to mount when mounting
the NFS-exported file system on the client system.
There are three ways to configure an NFS file system export. On demand via the command line (client side), automatically via the /etc/fstab file (client side), and automatically via autofs configuration files, such
as /etc/auto.master and /etc/auto.misc (server side with NIS).
For example, on demand via the command line (client side):
mount -o udp shadowman.example.com:/misc/export /misc/local
When the NFS mount is specified in /etc/fstab (client side):
server:/usr/local/pub /pub nfs rsize=8192,wsize=8192,timeo=14,intr,udp
When the NFS mount is specified in an autofs configuration file for a NIS server, available for NIS enabled workstations:
myproject -rw,soft,intr,rsize=8192,wsize=8192,udp penguin.example.net:/proj52
Since the default is TCP, if the -o udp option is not specified, the NFS-exported file system is accessed via TCP.
The advantages of using TCP include the following:
-
Improved connection durability, thus less NFS stale file handles messages.
-
Performance gain on heavily loaded networks because TCP acknowledges every packet, unlike UDP which only acknowledges completion.
-
TCP has better congestion control than UDP. On a very congested network, UDP packets are the first packets that are dropped. This means that if NFS is writing data (in 8K chunks) all of that 8K must be retransmitted over UDP. Because of TCP's reliability, only parts of that 8K data are transmitted at a time.
-
Error detection. When a TCP connection breaks (due to the server being unavailable) the client stops sending data and restarts the connection process once the server becomes available. With UDP, since it's connection-less, the client continues to pound the network with data until the server reestablishes a connection.
The main disadvantage is that there is a very small performance hit due to the overhead associated with the TCP protocol.
