Fragmentation

The IP-level datagram must be encapsulated in a lower network level packet
to travel in the network. The ideal case could be considered to be that in
which a network header is added to the datagram and is sent through the
same network. In the real case, the size of the packet of the network can vary
greatly and can be smaller than the size of the IP datagram, so it is necessary
to fragment the IP datagram in several pieces to send it through the network.
This is where it is necessary to introduce the concept of maximum transfer
unit (MTU), which is the maximum size of a packet that can be transferred
by a specific network.
To approach this problem, the size of the IP datagram could have been
limited to the smallest MTU of the networks, but this system would be
highly inefficient in networks where a larger sized packet were admitted.
Really, the solution consists in, on the one hand, choosing a suitable
packet size, and, on the other, designing how to divide the datagram in fragments
when it is necessary.
The gateways undertake this fragmentation, and the rules for the fragmentation
are as follows:
• The size of the resulting fragments must be a multiple of an octet so
that the data displacement records, offset, within the datagram are
done correctly.
• The sizes of the fragments are freely chosen (i.e., the initial size does
not have an influence and the resulting fragments can be of differing
lengths).
• The gateways must accept datagrams with a greater size than that of
the network they are connected to. This is so larger datagrams can be
admitted to the network, although they must be fragmented to
travel in it.
• The hosts and gateways must handle datagrams larger than 576
octets. Each fragment has the same format as the original datagram, which
means that the fragments generated will have a header that is the same as the
original datagram, except for a series of differences explained next. The data
of each fragment will be made up of fragments of the original datagram’s
data. The size of each fragment will be limited by the MTU of the network
where they are going to be sent.
We will now look at the differences between the header of the original
datagram and those of its fragments. If we consider that the original datagram
has been divided into N fragments, the first N−1 fragments will have
the MF flag set to 1 and the fragment N will have this flag set to 0, indicating
that there are no more fragments. The fragment offset field is used to indicate,
when the datagram is a fragment, which position in the original datagram
the data are located. The fragmentation is always done in multiples of
eight bytes, which is the elemental unit of fragmentation. Thanks to this, the
field requires only 13 bits instead of the 16 that would be necessary if bytes
were counted. Of the 3 bits saved, two are used in the DF and MF flags and
the third is not used. Note that the entity that must reassemble the datagram
would have no way of knowing its total length if it were not for the last fragment
with its fragment offset and total length fields, which permit it to know
the total length of the datagram. In both cases, the identification field, along
with the datagram source address, allow the gateway/destination to know
which arriving fragments belong to which datagrams.
At this stage we ask the following question: Who must reassemble the
fragmented datagram? There are two options: The first is that the first gateway
receives the fragments at the next routing point of the Internet; the second
option is that the reassembly is done by the destination host of the
original datagram.
The disadvantages of the second option, which is implemented in practice,
are first that the fragments can cross networks with a greater MTU without
taking advantage of its capacity to transmit larger sized packets. Second, it
is obvious that if one fragment of a datagram is lost, the whole datagram is lost.
In spite of this, the system works well and has the advantage that the
intermediate gateways do not need to store and reassemble packets, which
means the gateways are simpler and require less capacity.
Finally, there is another interesting aspect to deal with related to the
way the MF bit is propagated if it is necessary to further fragment the datagram’s
fragments. It is done in a very simple way. For all of the fragments
with the MF bit at 1, all its fragments have this bit set at 1. On the contrary,
when a fragment is received with the MF bit at 0, this bit is set at 1 for all the
fragments except the last one, which is set to 0.Once the fragments arrive at a destination address, they must be reassembled.
For this, the IP software must have a system implemented such
that:
• It enables the rapid insertion of the fragment in the group. As the
fragment offset field of the fragments refers to the original datagram,
space can easily be reserved for each fragment.
• It must have an efficient test of the whole datagram.
• It must quickly eliminate the stored fragments if they surpass the
maximum time of reassembly.
If at a particular moment the fragment storage capacity is exceeded, all
of the fragments are discarded. This is done because one lost fragment makes
it impossible to obtain the whole datagram; in this way, space is liberated for
other datagrams to be reassembled.