Implementing a Large Non-Linear Integer Optimization on a Distributed Collection of Office Computers
Ronald H. Nickel, Igor Mikolic-Torreira, Jon W. Tolle · Fluids Engineering · 2000
Abstract We describe how we implemented the MIMERBS non-linear integer optimization methodology to run across a virtual super-computer of existing Windows NT computers networked together by an ordinary office LAN. We describe how we configured this virtual computer and how we parallelized MIMERBS to work efficiently in view of the high communications costs of our virtual computer. We also describe how we made MIMERBS highly fault-tolerant and dynamically configurable; in particular we describe techniques for handling the loss of individual computers, for automatic on-the-fly addition of new computers, and for dynamic load-balancing. We also describe the techniques we used to share computer resources gracefully with officer workers using the same computers concurrently for ordinary word and data processing. We present performance results from specific MIMERBS applications. These examples show that performance of several gigaFLOPS is possible with just a few dozen ordinary computers on an office LAN.