| Modularity is an essential element of scalable engineering; in comparison to over-optimized “one-off” designs, modular designs are invariably more scalable, readily maintainable and significantly more reliable. However, modular and layered designs often are accompanied by an increased price, typically due to added inefficiency, in the resulting implementation.; In this dissertation, we study some of the inefficiencies introduced by conventional layered and modular architectures. This dissertation makes a contribution in the finding that the antidote to some of the causes of inefficiencies can be surprisingly simple. We introduce a methodological approach, called Dynamic Composability, which can help build composite systems that are either significantly more efficient and compact or are more agile and flexible in their responsiveness to the environment. A significant amount of optimization of the system can be obtained at system composition time, by following a systematic methodology which is well balanced between component implementation and system composition. A further contribution is the demonstration that these performance and cost improvements can be obtained without sacrificing the tangible engineering-process related advantages of modular and layered design, namely the advantages of scalability, maintainability and reliability.; In particular, in the case of networking and operating systems software, we demonstrate methodology that can be adopted before and during the implementation of the components that make up the system, and corresponding procedures that can be adopted during offline system composition and during on-line dynamic composition. We demonstrate this by constructing three compelling and concrete implementations—a networking stack which is three orders of magnitude more compact than conventional stacks (demonstrating gain in compactness), an architecture for ad-hoc networking among dynamic and mobile nodes (demonstrating gains due to adaptiveness) and a solution for adaptive overload management in AWACS radar software (demonstrating real-time and fault-tolerance properties). |
