Computational Models for Information Reuse

Abstract

Telecom Italia, Divisione Ricerca e Sviluppo, ItalyEmail: arcelli@dia.unisa.itTechniques that optimize computations by reusing partial results have a long tradition in computerscience. Seen from the point of view of sequential computations, all these techniques share thecommon execution strategy of storingpartial results ina centralized datastructure, whiletheydifferas to how the results are computed, e.g. in a data-driven or constraint-driven fashion. Concurrentsystems, namely the wide variety of systems that range from fine-grained parallelism to coarse-grained distribution, add another variable into the game. In fact, information reuse is here tangledwith issues of local memory of agents and inter-agent communication. Thus, optimal strategies forinformation reuse directly affect agent configurations as well as agent communication protocols andstrictly depend on those morphological aspects of the computational domain related to the sharingof structures among different data values. In this paper, we define a formal framework suitable forthe study of information reuse from the point of view of concurrent systems. The main result ofour work is in the identification of two distinct morphological features of computational domains,namely recursively replicated structures and structure copying. These features induce two differentforms of information reuse that can be optimized, respectively, by solipsistic agents with large localmemory and by large bandwidth networks of collaborative agents.Received April 21, 1998; revised September 14, 1999