+.SS "Using NFS over UDP on high-speed links"
+Using NFS over UDP on high-speed links such as Gigabit
+.BR "can cause silent data corruption" .
+.P
+The problem can be triggered at high loads, and is caused by problems in
+IP fragment reassembly. NFS read and writes typically transmit UDP packets
+of 4 Kilobytes or more, which have to be broken up into several fragments
+in order to be sent over the Ethernet link, which limits packets to 1500
+bytes by default. This process happens at the IP network layer and is
+called fragmentation.
+.P
+In order to identify fragments that belong together, IP assigns a 16bit
+.I IP ID
+value to each packet; fragments generated from the same UDP packet
+will have the same IP ID. The receiving system will collect these
+fragments and combine them to form the original UDP packet. This process
+is called reassembly. The default timeout for packet reassembly is
+30 seconds; if the network stack does not receive all fragments of
+a given packet within this interval, it assumes the missing fragment(s)
+got lost and discards those it already received.
+.P
+The problem this creates over high-speed links is that it is possible
+to send more than 65536 packets within 30 seconds. In fact, with
+heavy NFS traffic one can observe that the IP IDs repeat after about
+5 seconds.
+.P
+This has serious effects on reassembly: if one fragment gets lost,
+another fragment
+.I from a different packet
+but with the
+.I same IP ID
+will arrive within the 30 second timeout, and the network stack will
+combine these fragments to form a new packet. Most of the time, network
+layers above IP will detect this mismatched reassembly - in the case
+of UDP, the UDP checksum, which is a 16 bit checksum over the entire
+packet payload, will usually not match, and UDP will discard the
+bad packet.
+.P
+However, the UDP checksum is 16 bit only, so there is a chance of 1 in
+65536 that it will match even if the packet payload is completely
+random (which very often isn't the case). If that is the case,
+silent data corruption will occur.
+.P
+This potential should be taken seriously, at least on Gigabit
+Ethernet.
+Network speeds of 100Mbit/s should be considered less
+problematic, because with most traffic patterns IP ID wrap around
+will take much longer than 30 seconds.
+.P
+It is therefore strongly recommended to use
+.BR "NFS over TCP where possible" ,
+since TCP does not perform fragmentation.
+.P
+If you absolutely have to use NFS over UDP over Gigabit Ethernet,
+some steps can be taken to mitigate the problem and reduce the
+probability of corruption:
+.TP +1.5i
+.I Jumbo frames:
+Many Gigabit network cards are capable of transmitting
+frames bigger than the 1500 byte limit of traditional Ethernet, typically
+9000 bytes. Using jumbo frames of 9000 bytes will allow you to run NFS over
+UDP at a page size of 8K without fragmentation. Of course, this is
+only feasible if all involved stations support jumbo frames.
+.IP
+To enable a machine to send jumbo frames on cards that support it,
+it is sufficient to configure the interface for a MTU value of 9000.
+.TP +1.5i
+.I Lower reassembly timeout:
+By lowering this timeout below the time it takes the IP ID counter
+to wrap around, incorrect reassembly of fragments can be prevented
+as well. To do so, simply write the new timeout value (in seconds)
+to the file
+.BR /proc/sys/net/ipv4/ipfrag_time .
+.IP
+A value of 2 seconds will greatly reduce the probability of IPID clashes on
+a single Gigabit link, while still allowing for a reasonable timeout
+when receiving fragmented traffic from distant peers.
projects / nfs-utils.git / blobdiff