The Development of a Grid Based Engineering Design Problem Solving Environment

To enhance engineering insight, reduce development costs and improve product quality, design studies are increasingly using sophisticated analysis packages together with optimisation tools. This combined approach requires a high level of systems integration and the ability to exploit clusters of networked workstations to achieve efficient and re-usable design systems. What is a PSE " A PSE is a computer system that provides all the computational facilities needed to solve a target class of problems. " (J. Rice-Purdue University) PSE Components The main PSE subsystems include visual-programming environments for graphically composing, steering and form a task graph. This defines the execution order and dependencies between the various components making up a job. For a component to be available for use in VCCE, an XML definition of the component must be available in the Component repository. In the Design of Experiments(DoE)/Response Surface Modelling(RSM) process (left part of diagram above) some technique (e.g., Latin hypercubes) is used to select a set of initial design points (typically say 100 points) that will span the design space. The analysis codes are then run for each design point, ideally in parallel on some large compute cluster, as the design point evaluations are independent of each other. From the resulting database, a response surface is built using the chosen approximation method (e.g., kriging). If satisfactory, the surface is searched for an optimal design point (right part of diagram) and if found, full analysis codes are run at that point. If the analysis is satisfactory and all design constraints are met, the process terminates; otherwise the new design point is added to the design set, the surface regenerated and the optimisation repeated. From a user perspective, the RSM process suggests a data-centric approach to the computational workflow, where the user is primarily looking at problem design point and results data to-date (and possibly other test data such as flight trial or wind tunnel results in aerospace design problems) in order to decide where new problem analyses need to be run. From a computational perspective, the aim is to achieve the maximum amount of parallelism in the analysis computations by exploiting the various cluster resources available. This can be accomplished by de-coupling the search engine computation from the analysis computations and using some form of task farming to schedule parallel running of analysis tasks over the available resources. Rapid advances in commodity computing and the emerging nationwide computational grid …