The project builds on a recently completed NSF project, in which the PI developed a prototype of a dynamically adaptive compilation system. This system has demonstrated significant gains in overall program performance in science and engineering applications. There is tremendous potential in moving compiler optimization decisions to runtime, where complete information is available. For example, knowing the input to a code section would allow the compiler to directly produce this code's output, delivering the program response instantly. The new concepts and implementations will (i) enable a new generation of applications that improve as they age and adapt to changes unforeseen by their developers, (ii) provide solutions to key open issues in the design of current compilers, such as where and when to apply what optimization techniques, (iii) address new problems arising in tomorrow's increasingly dynamic, mobile, and unreliable execution environments.
Broader Impact:
While the narrow view of the project goals is to develop advanced compiler techniques for high performance computer applications, the successful completion of the proposed work will have a broader impact. These techniques will be distributed to publicly available compiler infrastructures using the Polaris and Cetus compilers. The PI has been distributing the Polaris compiler to other researchers for many years, and it is one of the most advanced and widely known HPC compiler research infrastructures. These infrastructures have been used in courses among Students using the newest compiler technology. Using this basis, we will make a concerted effort to combine the proposed research with educational activities at the undergraduate and graduate level. Furthermore, the proposed effort addresses what is Believed to be the most severe limitation of the current generation of compilers, which is the compilers' conservative assumptions, due to the lack of runtime knowledge. Dynamically adaptive techniques have the potential to increase the power of compiler optimizations very significantly and hence contribute to improved programmability of high-performance computing applications in general. Finally, while our goal is to advance those application areas where performance is the most critical value -high-performance computing applications -our techniques are generic in nature and will thus benefit all classes of software.
