We propose an approximate model for a nonsaturated IEEE 802.11 DCF network. This model captures the significant influence of an arbitrary node transmit buffer size on the network performance. We find that increasing the buffer size can improve the throughput slightly but can lead to a dramatic increase in the packet delay without necessarily a corresponding reduction in the packet loss rate. This result suggests that there may be little benefit in provisioning very large buffers, even for loss-sensitive applications. Our model outperforms prior models in terms of simplicity, computation speed, and accuracy. The simplicity stems from using a renewal theory approach for the collision probability instead of the usual multidimensional Markov chain, and it makes our model easier to understand, manipulate and extend; for instance, we are able to use our model to investigate the important problem of convergence of the collision probability calculation. The remarkable improvement in the computation speed is due to the use of an efficient numerical transform inversion algorithm to invert generating functions of key parameters of the model. The accuracy is due to a carefully constructed model for the service time distribution. We verify our model using ns-2 simulation and show that our analytical results based on an M/G/1/K queuing model are able to accurately predict a wide range of performance metrics, including the packet loss rate and the waiting time distribution. In contradiction to claims by other authors, we show that 1) a nonsaturated DCF model like ours that makes use of decoupling assumptions for the collision probability and queuing dynamics can produce accurate predictions of metrics other than just the throughput, and 2) the actual service time and waiting time distributions for DCF networks have truncated heavy-tailed shapes (i.e., appear initially straight on a log-log plot) rather than exponential shapes. Our work will help developers select appropriate buffe- - r sizes for 802.11 devices, and will help system administrators predict the performance of applications.