Peer-to-peer protocols play an increasingly instrumental role in Internet content distribution. Consequently, it is important to gain a full understanding of how these protocols behave in practice and how their parameters impact overall performance. We present the first experimental investigation of the peer selection strategy of the popular BitTorrent protocol in an instrumented private torrent. By observing the decisions of more than 40 nodes, we validate three BitTorrent properties that, though widely believed to hold, have not been demonstrated experimentally. These include the clustering of similar-bandwidth peers, the effectiveness of BitTorrent's sharing incentives, and the peers' high average upload utilization. In addition, our results show that BitTorrent's new choking algorithm in seed state provides uniform service to all peers, and that an underprovisioned initial seed leads to the absence of peer clustering and less effective sharing incentives. Based on our observations, we provide guidelines for seed provisioning by content providers, and discuss a tracker protocol extension that addresses an identified limitation of the protocol.