Many works have studied the performance of TCP by modeling the network as a single bottleneck node, the one having the slowest outgoing rate on the path. In these works, the behavior of the protocol, especially during slow-start, has been characterized as a function of this main bottlenec- k. In this paper, we present a more general model taking into account all the nodes on the path. We show that, in addition to the main bottleneck, the performance of the protocol can be seriously affected by the parameters of the other nodes. The other nodes may, in some cases, cause an improvement in the performance since they may decrease the burstiness of traffic arriving at the main bottleneck. They could, on the other hand, cause performance degradation in case losses occur in those nodes. Indeed, if the buffers in these nodes are not well dimensioned, then the congestion can be shifted to them even if their outgoing rate is faster than that of the main bottleneck- . The effect of the other nodes on the behavior of TCP during Congestion Avoidance can be neglected. However, we show analytically that the slow-start behavior predicted by a single node model can be completely different when considering the general model. Our analysis is followed by guidelines for the dimensioning of network buffers so as to improve the performance of TCP. Also, as a result of this general model, we are able to identify, understand and analyze a problem in the operation of TCP which results in a throughput deterioration and which cannot be explained by a single node model. By simulations, we show that this problem can be resolved if RED (Random Early Detection) buffer management policies are deployed in the Internet.