Abstract (0905368) Many important scientific computing problems, called ?streaming applications,? have high input data rates derived from real-time sensor data or directly from data streaming from disk arrays. Real-time sensor based data (e.g., telescopic astrophysical data obtained in the search for new planets) is frequently sourced from analog devices and requires filtering and various data cleaning prior to performing a host of complex computations. Large disk based data sets (e.g., genome and protein sets used in understanding disease factors) are often passed at high data rates from disk storage. Choices for dealing with such applications include a multiplicity of computing devices (e.g., general purpose processors, chip-multiprocessors, graphics processors, field programmable gate arrays, etc.). While each individually is well matched to certain types of computations, often more effective solutions are found by integrating multiple computer types into a single system. The central research issue is determining how to effectively integrate diverse computing resources for solution of complex streaming applications. The research includes further development of the AutoPipe design environment. AutoPipe provides tools for algorithm specification, and for design, simulation and deployment of diverse integrated computing architectures. Techniques for inclusion of analog devices in mixed analog-digital systems is being undertaken so that Auto-Pipe can handle mixed signal, analog/digital algorithmic functional and resource components in a single system. The research activity is driven by two important applications taken from the astro-physics and computational biology domains. There will be heavy involvement of graduate and undergraduate students in the research.
This research has focused on improving the performance of scientific computing problems characterized as streaming applications. Such applications often have high input data rates derived from real-time sensor data or from unstructured data files streaming from disk arrays. We have designed and built the Auto-Pipe application development environment, which enables the use of hybrid computer systems composed of traditional multicore processors, field-programmable gate arrays (FPGAs), and graphics processing units (GPUs). In the area of general streaming applications, we have developed a theory of deadlock avoidance in the presence of filtering kernels, including a proof of the necessary and sufficient conditions for deadlock, a pair of deadlock avoidance algorithms, and computation of the optimal parameterization of the tradeoff between extra communication overheads and buffering requirements. In addition, we have developed a new asymptotic model of computation on GPUs that effectively considers the unique properties of these architectures, exploiting thread-based computation to hide latency to global memory. In the area of computational biology, we have developed a new algorithm for short read alignment, the comparison of short strings of DNA to a larger reference database. And in the area of polarization-enabled optics, we have constructed a 4 mega pixel imaging sensor that has important applications to tumor imaging in cancer patients. The high spatial resolution of this imager is also important as a key aspect and requirement for astrophysics imaging applications.
