Optical fibre communications are now widely used in many applications, including local area computer networks. I postulate that many future optical LANs will be required to operate with limited optical power budgets for a variety of reasons, including increased system complexity and link speed, low cost components and minimal increases in transmit power. Some developers will wish to run links with reduced power budget margins, and the received data in these systems will be more susceptible to errors than has been the case previously.

The errors observed in optical systems are investigated using the particular case of Gigabit Ethernet on fibre as an example. Gigabit Ethernet is one of three popular optical local area interconnects which use 8B/10B line coding, along with Fibre Channel and Infiniband, and is widely deployed. This line encoding is also used by packet switched optical LANs currently under development. A probabilistic analysis follows the effects of a single channel error in a frame, through the line coding scheme and the MAC layer frame error detection mechanisms. Empirical data is used to enhance this original analysis, making it directly relevant to deployed systems.

Experiments using Gigabit Ethernet on fibre with reduced power levels at the receiver to simulate the effect of limited power margins are described. It is found that channel bit error rate and packet loss rate have only a weakly deterministic relationship, due to interactions between a number of non-uniform error characteristics at various network sub-layers. Some data payloads suffer from high bit error rates and low packet loss rates, compared to others with lower bit error rates and yet higher packet losses. Experiments using real Internet traffic contribute to the development of a novel model linking packet loss, the payload damage rate, and channel bit error rate. The observed error behaviours at various points in the physical and data link layers are detailed. These include data-dependent channel errors; this error hot- spotting is in contrast to the failure modes observed in a copper-based system. It is also found that both multiple channel errors within a single code-group, and multiple error instances within a frame, occur more frequently than might be expected. The overall effects of these error characteristics on the ability of cyclic redundancy checks (CRCs) to detect errors, and on the performance of higher layers in the network, is considered.

This dissertation contributes to the discussion of layer interactions, which may lead to un-foreseen performance issues at higher levels of the network stack, and extends it by considering the physical and data link layers for a common form of optical link. The increased risk of errors in future optical networks, and my findings for 8B/10B encoded optical links, demonstrate the need for a cross-layer understanding of error characteristics in such systems. The development of these new networks should take error performance into account in light of the particular requirements of the application in question.