A success story witnessing the validity of the GISELA WP3 methodology is the porting of Industry@Grid applicationon the GISELA Grid infrastructure. In the context of a Memorandum of Understanding between EPIKH and GISELA, Luiz Rossi, from the Centro Federal de Educação Tecnológica Celso Suckow da Fonseca (CEFET-RJ) of Rio de Janeiro (Brazil), has spent one month at INFN Catania to learn Grid technology and to "gridify" the Industry@Grid application. At the end of this period the application was successfully ported to the Latin American Grid and Luiz came back to Brazil with the expertise to spread the Grid paradigm within his organization and country.
The Industry@Grid ShopScheduling application uses LiSA (http://lisa.math.uni-magdeburg.de/) routines to solve scheduling problems representing decisions that manufacturing industries face on daily or weekly basis. Business restrictions such as due dates, machinery availability, operations precedence's among other, can be modeled to represent factory floors. The application solves the scheduling problem to help the decision makers to plan the production in order to respond to market demands on time, trying to keep costs down so the company can generate bigger profits or charge cheaper prices. These decisions are hard to take due to the computational complexity to solve this kind of problems to optimality. Although many different heuristics available generate good solutions, they may not be close to the absolute maximum in the space of parameters. Neighborhood searches may find local maxima and stop execution while significantly better solutions may exist. The use of grid computing allows researchers and practitioners to experiment and run the application millions of times that otherwise could take months running on a single desktop machine. One of the latest experiments was run with a sample of one million problems. It was designed to run on 150 CPU cores, with an average solution time of 11 CPU hours. The total computational time was 1650 hours (more than 2 months). Although the results were not optimal, they are close to the benchmark solution, yielding considerably good results in an affordable amount of time for the industry. The application is designed to work in parallel solving millions of instances of the same problem but in a different fashion. From an initial problem, the application is capable of generating random initial solutions and to apply neighborhood search algorithms to improve each different solution. After execution, the results are filtered and only the better solution is brought back to the user (and a log reporting all the objectives found, so a graphic can be drawn and different analysis made). The dissemination of this approach may allow not only better solutions for companies' shop scheduling operations, but also substantial economies in hardware and software investments.