Towards a compilation paradigm for computational applications on the information power grid

The Information Power Grid (IPG) refers to the world-wide infrastructure of computers and their interconnections. We are only at the very beginning of exploring applications and system software that can exploit these resources effectively and, at the same time, provide adequate interfaces to the programmer and end-user. In this paper, we discuss compiler technology that serves this purpose. The core consists of methods and services that enable an application to adapt flexibly to the dynamically changing nature of the IPG. We present several applications that demonstrate performance opportunities in such scenarios. We then present and evaluate new compiler techniques and services that allow an application to be dynamically re-optimized as the execution environments change. One very important property of such system support is that the overheads of dynamic re-optimizations are kept small. We will show that this is the case for the presented techniques. The resulting applications can run close to the best performance that could be obtained with prior knowledge of the runtime situations.