The goal of meta-computing is to harness diverse high performance computing resources into one integrated computing environment which allows the transparent remote execution of applications, makes use of exiting parallel and distributed resource management facilities, and allows wide-area parallel processing. Enabling transparent execution without a significant loss in performance requires new solutions to arising communication and synchronization problems.
In this project we focus on three major areas relevant to meta-computing: termination detection, data-driven communication, and support for spatial data sets. Termination detection is a crucial synchronization primitive in data-driven applications. We study termination detection techniques tailored towards meta-computing characteristics and identify application scenarios which do not require the full power of generic termination detection. Our goal is to develop a framework for incorporating new termination detection primitives into run-time systems of message-passing and multithreaded environments. Our work on communication operations concentrates on data-driven, irregular communication primitives in message-passing and multithreaded systems; in particular data-driven group communication. Our work will demonstrate that portable and scalable data-driven communication can be embodied into existing communication interfaces such as MPI. The importance of effectively handling massive spatial data sets in meta-computing environments is well-documented. We are investigating on how to use hierarchical spatial representations as the basis for new distributed data structures suitable for meta-computing environments. We are developing load balancing methods which make use of the underlying hierarchical structure and we study the effect of data migration on the overall performance.
For additional information see http://ece.www.ecn.purdue.edu/~scalable/.