Pre-Emptive Network Management

Abstract

Instead of waiting for systems to fail before taking remedial action, it might be possible to remove systems from service by diverting traffic prior to failure. Also, the likely failure mechanism might be identified from the transient pattern of error events, facilitating rapid repair and restoration. If these objectives were achieved they would lead to extensive cost savings and network performance enhancement. The basis for this alternative monitoring strategy has been formulated from practical system experience and predicated by the total lack of adequate burst error models. To date it has been assumed that randomly generated errors have a negative exponential arrival statistic, which can be demonstrated to be true. All other errors have been assumed to have some form of compound Poisson arrival [1]. Whilst this might be true, no one has been successful in formulating a general mathematical model that fits anything but a small selection of the recorded error events from practical networks. There are two possible explanations for this apparent difficulty. First the statistics of individual bursts may be different depending on their origin, such as power transients, lightning, capacitor breakdown, human intervention, radio interference, etc. Secondly, the reported models generally try to fit statistical distributions to the error signals produced after the line decoding operation. It is contended here that decoding and retiming circuits significantly distort the burst error statistics and further complicate the model.