Be mindful that fragmentation may cause low link utilization. In other words, during fragmentation a node may end up transporting mainly overhead bytes in the fragment as opposed to payload bytes. This would only intensify the problem that fragmentation is intended to solve, especially on an ATM access link that tend to carry larger encapsulation overhead.
To reduce the overhead associated with fragmentation, the following is enforced in the 7750 SR:
The minimum fragment payload size is at least 10times greater than the overhead (MLPPP header, ATM Encapsulation and AAL5 trailer) associated with the fragment.
The optimal fragment length (including the MLPPP header, the ATM Encapsulation and the AAL5 trailer) is a multiple of 48B. Otherwise, the AAL5 layer would add an additional 48B boundary padding to each fragment which would unnecessarily expand the overhead associated with fragmentation. By aligning all-but-last fragments to a 48B boundary, only the last fragment potentially contains the AAL5 48B boundary padding which is no different from a non-fragmented packet. All fragments, except for the last fragment, are referred to as non-padded fragments. The last fragment is padded if it is not already natively aligned to a 48B boundary.
As an example, calculate the optimal fragment size based on the encapsulation criteria with the maximum fragment overhead of 22B. To achieve >10x transmission efficiency the fragment payload size must be 220B (10*22B). To avoid the AAL5 padding, the entire fragment (overhead + payload) is rounded UP on a 48B boundary. The final fragment size is 288B [22B + 22B*10 + 48B_allignment].
In conclusion, an optimal fragment size was selected that carries the payload with at least 90% efficiency. The last fragment of the packet cannot be artificially aligned on a 48B boundary (it is a natural reminder), so it is be padded by the AAL5 layer. Therefore, the efficiency of the last fragment is less than 90% in the example. In the extreme case, the efficiency of this last fragment may be only 2%.
For the Ethernet-based last mile, the CPM always makes sure that the fragment size plus encapsulation overhead is larger or equal to the minimum Ethernet packet length of 64B.