BitTorrent is a very scalable _le sharing protocol that utilizes the upload bandwidth of peers to of_oad the original content source. With BitTorrent, each _le is split into many small pieces, each of which may be downloaded from different peers. While BitTorrent allows peers to effectively share pieces in systems with suf_cient participating peers, the performance can degrade if participation decreases. Using measurements of over 700 trackers, which collectively maintain state information of a combined total of 2.8 million unique torrents, we identify many torrents for which the system performance can be signi_cantly improved by re-allocating peers among the trackers. We propose a light-weight distributed swarm management algorithm that manages the peer torrents while ensuring load fairness among the trackers. The algorithm achieves much of its performance improvements by identifying and merging small swarms, for which the performance is more sensitive to _uctuations in the peer participation, and allows load sharing for large torrents.