System-level optimization of architectural performance under varying service demands

This paper describes the motivation, methods, and application of a regimen for optimizing cost-performance measures for distributed system architectures. Dynamic performance under varying workloads was optimized in terms of system resource parameters in an absolute time base context. The intent was to establish an exacting yet practical means for setting architecture parameters and tolerances near the outset of system-level development. An existing software-implemented prototype for an on-line transaction processing system (OLTPS) was modified and interfaced with a genetic algorithm (GA). An objective function composed of architecture attributes was used as a fitness metric for the GA. Five architecture parameters encoded in a 24-bit chromosome were manipulated by the GA to furnish instantiation values for the OLTPS prototype. After each prototype execution, four dynamic performance measures were fed to the GA for fitness ranking. Despite some hesitancy in GA convergence, the overall optimization scheme worked well. To achieve more rapid resolution of optima, a multiple mutation operator mechanism was improvised and used to advantage. In all, the effort expended for dynamically quantified system-level optimization is seen as well justified and quite beneficial to overall system design.