This paper entails an in-depth study of the system buffer size requirements for parallel scientific applications that communicate regularly with four or six neighbors. The goal of this study is to address two major issues:(1) for future system designs, which system features significantly impact the system buffer size requirements and (2) for systems where users can set the system buffer size (eg, IBM ËP2 and ËGI-Cray T3E), what is a good estimate to use for good performance. We propose an analytical model of buffer size requirements for message passing systems. This model is validated experimentally for the IBM ËP2 system. We then use the model to perform sensitivity analysis on buffer size requirements with varying network bandwidth, receive latency, memory copy rate, delay between messages, and different number of messages. The analytical results indicate that the two major factors that affect the buffer size are the delay between messages and the increase in network bandwidth. Further, these two factors are related. The network bandwidth has little impact on buffer size requirements when the delay is small, but has significant impact when the delay is large. The receive startup latency and receive memory copy rate have very little impact on buffer size requirements.