Reconfiguration for Fault Tolerance and Performance Analysis

Architecture reconfiguration, the ability of a system to alter the active interconnection among modules, has a history of different purposes and strategies. Its purposes develop from the relatively simple desire to formalize procedures that all processes have in common to reconfiguration for the improvement of fault-tolerance, to reconfiguration for performance enhancement, either through the simple maximizing of system use or by sophisticated notions of wedding topology to the specific needs of a given process. Strategies range from straightforward redundancy by means of an identical backup system to intricate structures employing multistage interconnection networks. The present discussion surveys the more important contributions to developments in reconfigurable architecture. The strategy here is in a sense to approach the field from an historical perspective, with the goal of developing a more coherent theory of reconfiguration. First, the Turing and von Neumann machines are discussed from the perspective of system reconfiguration, and it is seen that this early important theoretical work contains little that anticipates reconfiguration. Then some early developments in reconfiguration are analyzed, including the work of Estrin and associates on the fixed plus variable restructurable computer system, the attempt to theorize about configurable computers by Miller and Cocke, and the work of Reddi and Feustel on their restructable computer system. The discussion then focuses on the most sustained systems for fault tolerance and performance enhancement that have been proposed. An attempt will be made to define fault tolerance and to investigate some of the strategies used to achieve it. By investigating four different systems, the Tandern computer, the C.vmp system, the Extra Stage Cube, and the Gamma network, the move from dynamic redundancy to reconfiguration is observed. Then reconfiguration for performance enhancement is discussed. A survey of some proposals is attempted, then the discussion focuses on the most sustained systems that have been proposed: PASM, the DC architecture, the Star local network, and the NYU Ultracomputer. The discussion is organized around a comparison of control, scheduling, communication, and network topology. Finally, comparisons are drawn between fault tolerance and performance enhancement, in order to clarify the notion of reconfiguration and to reveal the common ground of fault tolerance and performance enhancement as well as the areas in which they diverge. An attempt is made in the conclusion to derive from this survey and analysis some observations on the nature of reconfiguration, as well as some remarks on necessary further areas of research