Recent developments in supercomputing have brought us massively parallel machines. With the number of processors multiplying, the appetite for more powerful applications that can take advantage of these large scale platforms has never ceased growing. Modern parallel applications typically have complex structure and dynamic behavior. These applications are composed of multiple components and have interleaving concurrent control flows. The workload pattern of these applications shifts during execution, causing load imbalance at run-time. Programming productivity, or the effectiveness and efficiency of programming high performance applications for these parallel platforms, has become a challenging issue. One of the most important observations during our pursuit of high productivity with scalable performance for complex and dynamic parallel applications was that adaptive resource management can and should be automated. The PPL research group has developed an Adaptive Run-Time System (ARTS) and a parallel programming language called Charm++ for automatic resource management via migratable objects. There are two obstacles in our pursuit of high productivity with the ARTS. The first is effective expression of global view of control in complex parallel programs. Traditional paradigms such as MPI and Global Address Space (GAS) paradigms, although popular, suffer from a drawback in modularity. For applications with multiple modules, they do not allow the runtime control over resource management of individual modules. Although Charm++ provides resources management capabilities and logical separation of multiple modules, its object-based message-driven model tends to obscure the global flow of control. We will explore new approaches to describing the flow of control for complicated parallel applications. As a reference implementation, we introduce a language {\em Charisma} for expressing the global view of control that can take advantage of the ARTS. We carry out productivity and performance study of Charisma, with various examples and real-life applications. The second issue is to efficiently accommodate existing prevalent programming paradigms. Different programming models suit different types of algorithms and applications. Also the programmer proficiency and preference may result in the variety of choices of programming languages and models. In particular, there are already many parallel libraries and applications written with prevalent paradigms such as MPI. In this thesis, we explore research issues in providing adaptivity support for prevalent paradigms. We will evaluate important existing parallel programming languages and develop virtualization techniques that bring the benefits of the ARTS to applications written using them. As a concrete example, we evaluate our implementation of Adaptive MPI in the context of benchmarks and applications. As applications grow in size, their development will be carried out by different teams with different paradigms, to best accommodate the expertise of the programmers and the requirements of the different application components. These paradigms include the new paradigm as represented by Charisma's global description of control, as well as existing ones such as MPI, GAS and Charm++. Charm++'s adaptive run-time system is a good candidate for a common environment for these paradigms to interoperate, and this thesis demonstrates the effectiveness of our research work for interoperability across multiple paradigms. The ultimate goal is to unify these various aspects and support multiparadigm parallel programming on a common run-time system for next-generation parallel applications.
