Interfaces, modularity and path lengths: The costs of implementing OSI transport

Abstract

Performance of communication software is influenced both by the underlying protocol architecture and by implementation. The debate as to which of the two kinds of factors dominates the overall performance has been going on for many years. In this paper we shed light on this issue by reporting on detailed performance measurements of our experimental OSI transport system. We focus on the analysis and classification of the path lengths involved during data transfer, obtaining insight into the relative costs of protocol functions versus system-level design and implementation choices. Our main conclusion is that adherence to general-purpose protocol standards is not the true bottleneck. The desire to provide a flexible general-purpose portable implementation with open inter-layer interfaces causes most of the overhead.