Buffering issues in TCP-based services

Abstract

Client-server protocols such as HTTP over a TCP transport layer may be implemented in a relatively straightforward manner, as is often given in 'textbook' examples. Many of these examples do not scale well to a full implementation across anything but a localhost connection with minimal data payload. Furthermore, such implementations are invariably less than optimal in terms of performance, robustness to differing remote client configurations, processor time utilization, and bandwidth usage. It is suggested that the examples often given in texts are overly simplified, to the point of being misleading. In this note, we examine some of these issues, resulting from the implementation of HTTP clients and servers using Unix (Berkeley) sockets, Windows NT (Winsock), and Java. Issues addressed include efficient send/receive buffering to minimizing kernel and user data copying, and robustness to variability in record delimiters.