|IBM scientists (left to right)|
Alessandro Curioni, Yves Ineichen and Costas Bekas
We spoke to Yves about his research at IBM and the potential impact of his work.
Yves, congratulations on winning the PRACE Award 2012 for building a general-purpose framework for multi-objective optimizations. What interested you in pursuing this research?
Thank you so much. The Paul Scherrer Institute has a project called SwissFel to build a free-electron laser accelerator. It's an extremely complex machine with numerous parameters or quality objectives that are mutually conflicting. My work approaches this machine as a mathematical, multi-objective optimization model.
Can you explain that in layperson's terms?
Sure. We encounter multi-objective optimization issues all the time in our daily lives. Let's say you want to buy a car, and you want to get the most value for the least amount of money. This is an everyday situation where we make decisions based on several — often conflicting — parameters. That's what we call a multi-objective optimization issue. If money were no object, you could buy the most expensive car with all the features anyone could want. But most of us want to maximize what we call optimality. That's not so simple because, if you improve one objective, it's likely that the others will deteriorate. In a nutshell, our work is to develop computational simulations to model this optimization process. The problem is computationally intensive; to tackle it we developed massively parallel implementations, in which we can utilize hundreds of thousands of cores.
What potential applications does this modeling tool have?
Simply put, the modeling tool accommodates change. Instead of remodeling, say, a factory floor, which can involve a costly "what if" search for an optimal solution, a computer simulation allows for many solutions to be tested at a negligible fraction of the cost. You see, it really makes a difference not to treat individual issues separately but all together. This not only yields a more optimal solution, it helps us to understand better the problem at hand.
Clients at IBM's Industry Solutions Lab (ISL) want this technology. IBM Research – Zurich has hosted executives from such sectors as automotive, aviation, supply chain, city planning — they all face these issues.
It could be important for many future applications, such as in the healthcare industry and many more. It's general enough to be applicable elsewhere in the future as well, and that's the beauty of it. It requires massively parallel computer power to solve, and that's the research field in software systems that interest me the most. And it fits right in with IBM's Smarter Planet concept, so I've enjoyed a lot of support.
What do you think was the crucial aspect of your work that earned the PRACE prize?
The core of the achievement was taking the original software for the model of the accelerator at PSI and scaling it from 10 to 10,000 parallel processors to optimize the model of the accelerator. Of particular difficulty was to design and implement algorithms and tools that schedule and deploy all the very dynamic parts (optimizers, forward solvers, etc.) of the framework on the parallel machine, keeping the load at a maximum, while also maximizing the computational efficiencies of the all individual blocks. The flexibility of the Blue Gene platform was key in achieving this goal. These efforts are the heart of what is called massive parallelization. Finally, I need to stress that our optimizer framework works for any number of different models, and this versatility is what makes our achievement so noteworthy.
All the best for your future research.
Thanks, there's still plenty of work to be done.
See also the original press release on the PRACE 2012 award.