A Tool for Performance-Driven Design of Parallel Systems

Performance-driven design is systems design focused on performance prediction. It demands heavy use of performance models throughout the design process, and the consequent blurring of the distinction between the roles “performance analyst” and “systems architect”. Large parallel systems performance prediction places its own extreme demands upon performance models. No monolithic model, utilizing one solution technique, can meet these demands. Thus a set of submodels, utilizing different solution techniques, is usually employed. This paper presents as a case study the modeling of a performance-driven, large, parallel database machine design. We have found that a “pipelined” system of submodels appears suitable to both the requirements of the design process and modeling large parallel systems. Examples illustrating the use of the pipelined modeling tool in our performance-driven environment are given.