An empirical study of distributed application performance

A major reason for the rarity of distributed applications, despite the proliferation of networks, is the sensitivity of their performance to various aspects of the network environment. It is demonstrated that distributed applications can run faster than local ones, using common hardware. It is also shown that the primary factors affecting performance are, in approximate order of importance: speed of the user's workstation, speed of the remote host (if any), and the high-level (above the transport level) protocols used. In particular, the use of batching, pipelining, and structure in high-level protocols reduces the degradation often experienced between different bandwidth networks. Less significant, but still noticeable improvements result from proper design and implementation of the underlying transport protocols. Ultimately, with proper application of these techniques, network bandwidth is rendered virtually insignificant.