Performability Evaluation of CSMA/CD & CSMA/DCR Protocols Under Transient Fault Conditions

CSMA/CD networks have gained widespread acceptance for use in many applications. However, a barrier to their acceptance for industrial use has been the nondeterministic nature of the collision-resolution scheme used in the standard variants (Ethernet and IEEE 802.3) of the protocol. Alternative schemes, which have a deterministically bounded collision-resolution time, have been proposed. These deterministic schemes typically work by partitioning the network nodes into groups, which can transmit during particular time periods after a collision is detected. The limit, in which each node is in its own group, has been implemented, but has not been evaluated in the presence of transient faults to determine if it offers any appreciable advantage over the standard exponential binary backoff algorithm. This is important in an industrial environment, where noise can appreciably affect protocol operation. This paper presents the results of such an evaluation for CSMA/CD and a deterministic protocol under workloads anticipated in an industrial environment. The paper, 1) illustrates the utility of stochastic activity networks in representing both performance- and dependability-related behaviors in one model, and 2) provides appreciable insight into the performability of the two protocols. Modeling approach: Stochastic activity networks permit the representation of a relatively complex fault model as well as usual protocol operations. This is important when doing performability evaluations of this type. In addition, the fault model was built in a manner in which there was minimal interaction between it and the protocol models. The same fault model might thus be able to be used to investigate the performability of other protocols. Protocols: There are appreciable differences in the behavior of the deterministic and non-deterministic collision-resolution schemes, which depend on both the offered workload and rate at which transient faults occur. In the fault-free case, there is a distinct region of offered workload in which each resolution scheme performs best. Thus, if the workload were relatively constant and known, there would be situations when each resolution scheme should clearly be chosen. However, if the offered workload for the application is not known or varies appreciably, one should consider the absolute difference between the performance of the 2 collision-resolution schemes in each region. In this case, the non-deterministic scheme performs just slightly better than the deterministic scheme at low loads, while the deterministic scheme performs appreciably better at high loads. Thus the deterministic resolution scheme offers real advantages when the offered workload is high, is not known, or varies appreciably. Where transient fault conditions exist, the effects due to transient faults have an appreciable impact on the behavior of the non-deterministic collision-resolution scheme. The most pronounced effect is evident when the workload is heavy and the occurrence of transient faults is frequent. The non-deterministic collision-resolution protocol becomes unstable at 70% of full load, when many transient faults occur. On the other hand, the deterministic-collision resolution protocol continues to perform acceptably, although there is some degradation in performance. Thus when transient faults are considered, the deterministic collision resolution scheme clearly performs better than the non-deterministic scheme.