Avaliação do impacto de falhas no tráfego da rede da Universidade Federal de Juiz de Fora

Abstract

In this paper we characterize the impact of failures in the RNP´s (Rede Nacional de Ensino e Pesquisa) network on data traffic at UFJF (Universidade Federal de Juiz de Fora). In particular, we study the impact of failures on traffic, user, and application behavior. We classify the failures we study into partial failures and performance problems. Partial failures are interesting in that they persist for several hours and impact only international links, so destinations hosted in Brazil remain reachable. On the other hand, performance problems, caused by the disruption of connectivity between important links in the RNP backbone, without causing loss of end-to-end connectivity between the university and destinations in Brazil and abroad. Our results show that, during partial failures, although failures in international links have negligible impact on the performance of national traffic, users do adapt their behavior to the unavailability of services hosted abroad. For example, entertainment traffic migrates from Facebook to YouTube, which remains reachable during the analyzed failures; and the fraction of interactive traffic gradually decreases during failures, indicating that users may leave the campus early. We also show that, during partial failures, asynchronous applications hosted abroad, like Dropbox and SMTP, queue up tasks during the failure and cause traffic bursts when the failure is restored. We found that performance failures primarily impact the volume of data transmitted and metrics such as RTT. We have observed that when performance problems occurs, contrary to expectations, the performance improves for some destinations, due to changes in the routes used as primary path for transmissions. Our results can be used to guide improvements in the infrasctruture of the RNP backbone and assist in better defining parameters of route configurations to minimize the impact of some kind of failures.