(PhD08) Petalisp - Towards Fully Dynamic Program Optimization
TimeMonday, June 17th1pm - 6pm
DescriptionWe present a new technique for parallel programming, where computation is expressed by assembling inherently parallel data flow graphs at run time and by submitting them to an extremely fast evaluator. In doing so, we unlock a lot of potential for automatic parallelization and obtain a programming model that is elegant and productive. The downside of this approach is the added run time overhead of assembling, analyzing, optimizing and scheduling data flow graphs.
We provide a practical and mature implementation of this technique with the Common Lisp library Petalisp. Our early results are promising and show that Petalisp programs can often outperform equivalent NumPy code and are slowly getting competitive even to optimized C++ code.