Peer-to-peer systems contain a collection of equivalent entities that form an application-layer overlay network. Each entity acts as client, server, and overlay network router. Peer-to-peer systems enable the rapid development of distributed applications; in short order, peer-to-peer file sharing has become one of the most popular Internet applications. There are two primary functions in a peer-to-peer file sharing application: content discovery (determining which peers have the desired content) and content retrieval (transferring content from a peer to the requester). We focus on two aspects of content retrieval that have received limited attention, namely peer selection and request buffering. Using a combination of analytic modelling and simulation, we find that a simple greedy scheme that uses only local information to select a peer has performance comparable to alternative schemes, including those that make more globally-informed decisions. We also find that the optimal request buffering strategy depends on the system load and availability of peer information. For example, when peer information is scarce, download buffering works well; when peer information is plentiful, no buffering is preferable for better throughput. Our results have implications for the design of high-performance, low overhead peer-to-peer file sharing systems.