Mark what times you are available:
9:00am - 9:45am:
9:45am - 10:30am:
10:30am - 11:15am:
11:15am - noon:
1:30pm - 2:15pm:
2:15pm - 3:00pm:
3:00pm - 3:45pm:
3:45pm: talk setup
Hardware-conscious algorithm design for interactive physics simulation: Are theoretical research directions and software practices influenced by architectural traits of parallel platforms? (or can they afford not to?)
In this talk, the research area of interactive computational physics will be used as an example of how architectural intricacies often dictate software engineering practices, and sometimes even influence core theoretical developments. In recent years, the scope of computational physics has extended significantly beyond “traditional” High-Performance Computing application domains: Virtual models of living creatures and natural environments are a mainstay of motion pictures and entertainment-oriented computing, but are also finding an ever growing application base including medical diagnostics, surgical planning and virtual worlds. The performance potential of modern hardware has also inspired current and emerging applications to expand their demands beyond photorealistic rendering and visual detail. A number of areas will now associate visual fidelity and appeal with the ability of computer-generated models to resolve the anatomical function of virtual human bodies, or the intricate dynamics of natural phenomena. This talk will address research trends in computational physics for performance-sensitive visual computing applications, and how those relate to the architectural intricacies of commodity computing platforms.
Several case studies will be presented: Conventional examples will focus on fully featured modeling and simulation systems, which were developed as serial applications and parallelized using standard practices. On the opposite end of the spectrum, I will present certain research initiatives where all design decisions have been dictated by the pursuit of extreme performance and optimal platform utilization. Such approaches will be demonstrated to achieve orders of magnitude of performance in excess of conventional parallelization, albeit at a conscious expense of generality, versatility or coding convenience. Finally, I will pose the question whether extreme platform optimization will inevitably incur a sacrifice in generality or programming convenience, or if these two objectives can be reconciled.