Identifying and Modeling Flashcrowds in Global BitTorrent Systems

BitTorrent is currently one of the most popular peer-to-peer file sharing applications. Many previous studies of BitTorrent performance have focused on the dynamics of BitTorrent swarms and model flashcrowds by means of the arrival and departure processes of peers. However, until now there is no study that investigates the flashcrowd phenomenon at the BitTorrent system level. In this paper, we explore this phenomenon in global BitTorrent systems by analyzing three datasets collected from two BitTorrent trackers that serve millions of swarms. We first develop a flashcrowd identification algorithm that captures and quantitatively characterizes flashcrowds. Then, we conduct a comprehensive statistical analysis of the properties of flashcrowds, which include their magnitude, arrival time, duration, and plateau period. We also examine the overall arrival pattern of flashcrowds in BitTorrent systems, and the effects of flashcrowds on the growth and decay of BitTorrent systems. We formulate eleven findings that depict the big picture of the flashcrowds phenomenon in BitTorrent systems; for example, only a small fraction of swarms exhibit significant flashcrowds, but these swarms serve a large fraction of peers, from over a quarter to over two thirds of the peers during the peak period. We provide valuable insights for designing effective flashcrowd handling algorithms. We show that in BitTorrent flahscrowds can lead to extreme service deprivation, where the second seeder (the first fully served peer) only appears after the peak of the flashcrowd.