CARMA: Management infrastructure and middleware for multi-paradigm computing

A recent trend toward multi-paradigm systems that couple General-Purpose Processors (GPPs) and other special-purpose devices such as Field-Programmable Gate Arrays (FPGAs) has shown impressive speedups. In fact, some researchers see such systems as vital to extending an anticipated slowing of Moore's law. However, extracting the full potential of large-scale, multi-paradigm, High-Performance Computing (HPC) systems has proven difficult due in part to a lack of traditional job management services upon which HPC users have come to rely. In this dissertation, the Comprehensive Approach to Reconfigurable Management Architecture (CARMA) is introduced and investigated. CARMA attempts to address the issue of large-scale, multi-paradigm system utilization by (1) providing a scalable, modular framework standard upon which other scholarly work can build; (2) conducting feasibility and tradeoff analysis of key components to determine strategies for reducing overhead and increasing performance and functionality; and (3) deploying initial prototypes with baseline services to receive feedback from the research community. In order to serve the needs of a wide range of HPC users, versions of CARMA tailored to two unique processing paradigms have been deployed and studied on an embedded space system and Beowulf cluster.; Several scholarly contributions have been made by developing and analyzing the first comprehensive framework to date for job and resource management and services. CARMA is a first step toward developing an open standard with a modular design that will spur inventiveness and foster industry, government and scholarly collaboration much like Condor, Globus and the Open Systems Interconnect model have done for the cluster, grid and networking environments. Developing and evaluating components to fill out the CARMA infrastructure provided key insight into how future multi-paradigm runtime management services should be structured. In addition, deploying initial prototypes with baseline services will open multi-paradigm systems to a wider community of users, shorten the development learning curve, increase the utilization and usability of existing and future systems, reduce the cost of integrating future technology, and enable a wide community of users to share in future design and development efforts. CARMA will increase the usability of existing and future multi-paradigm systems while helping to shape future research.