Model-free Optimization

We consider how programming-in-the-large diiers from programming-in-the-small. We then focus on resolving a critical connict that arises when building large software systems. Large systems have both the requirement to be adequately eecient in execution and suuciently exible to be maintainable in a changing world. For our purpose their scale is assumed to be such that multiple processors, multiple applications, and multiple types of software modules will be simultaneously active. After a general background discussion, two critical points are elucidated: 1 The high level application process must obtain performance estimates as well as results from lower modules to create eecient schedules for each application type. 2 There will be no single optimization model or overall schedule, but this task will be as distributed and parallelized as the computations themselves. The general conclusion is that higher level abstractions are needed in all aspects of software technology to overcome the problems of programming-in-the-large and to use modern hardware technology eeectively. The 1992 DoD Software Technology Strategy (SWTS) emphasizes the importance of reuse of software. To achieve the goals`... a considerable challenge remains relating to validation and acceptance of assets developed outside the control of the user organization' SWTP:92, 4.1.4]. Techniques and standards for speciication, documentation, integration, and assurance have to become more eeective. Techniques which are successful in moderate scale software have not, or not yet, been able to demonstrate their eeectiveness in large-scale projects. Top-down decomposition fails when large systems must satisfy multiple, often connicting objectives. Careful system design prior to implementation fails when the time needed to complete the implementation exceeds the interval in which the environment and requirements change. Multi-layered approaches using the virtual machine paradigm show major performance deecits, so that the layers are often pierced in practice. Object-oriented programming has not yet proven scalability to mega-objects, probably because of the model problem, as elucidated through most of the remainder of this paper. There also remains the problem of mixed paradigms: 1 Objects as fully independent objects, accessible only via methods; an eeective view when modeling information processing systems : the smalltalk approach. 2 Objects as smart data structures, manipulable externally; an eeective view for the design community: the pdes approach.