ANDRADE, N. F.; http://lattes.cnpq.br/2729979018100977; ANDRADE, Nazareno Ferreira de.
Resumo:
The world’s largest encyclopedia, Wikipedia, a global telephony infrastructure, Skype, and one of the largest parallel computing platforms in the world, SETI@home, currently operate based on resources shared by those which, traditionally, were only clients of similar services. Sharing-based systems were propulsed by the prevalence of personal computers and Internet into major agents in our society. Understanding these systems and, from this understanding,
better design them, is therefore paramount for attaining the potential of this society. Understanding the behavior of a sharing-based system, however, is a complex enterprise. Besides understanding its design and implementation, it is also necessary to analyze the platform on which it runs and the workload it serves. Often, both this platform and workload result from the collective action of a large number of users, making their study non-trivial. In this work we examine these two aspects in a largely popular instance of sharingbased systems: peer-to-peer content distribution through BitTorrent. Our approach consists
of collecting and analyzing data about user behavior and system performance in multiple communities that rely on BitTorrent for content distribution, aiming to understand the impact of user behavior on the system’s performance. Our main results are (i) the understanding that the distribution of content popularity in BitTorrent communities is not long-tailed, what changes the trade-off in cache provisioning
for BitTorrent when compared to cache provisioning for the Web and for other peer-to-peer systems, (ii) a model for the decaying of the community’s interest for content items over time that is more precise than the state-of-the-art and contributes to more precise performance
evaluations of BitTorrent and similar systems, (iii) the analysis of user contribution levels in several communities that use BitTorrent, which shows that users contribute more in these communities than in other peer-to-peer systems and that users in a community contribute more in the presence of extra incentive mechanisms or when the community shares only strictly legal content and is socially cohesive, (iv) the conclusion that BitTorrent communities typically serve correctly virtually all requests received, and (v) the analysis of the efficacy of sharing enforcement mechanisms in promoting contribution in some BitTorrent
communities, which reveals that although these mechanisms result in equitable communities with higher levels of contribution, they do not lead to a larger proportion of requests served or to a noticeable change in the level of resource contention of the community when compared
to socially cohesive communities that share legal content.