This project entails the design, implementation and evaluation of computationally intensive signal processing applications on high-performance parallel embedded systems. As part of this project, we have developed and deployed techniques for parallelization, task mapping and allocation, parallel pipelined communication, data redistribution for applications consisting of several tasks. Another important goal of this project is to achieve a balance of throughput and latency by using the finite computational resources optimally.
We have completed an implementation of the Rome Labs PRI-Staggered Space-Time Adaptive Processing (STAP) application. This STAP algorithm involves (1) Doppler filter processing, (2) weight computation, (3) beam forming, (4) pulse compression, and (5) CFAR detection. We designed a parallel pipelined implementation. The implementation is portable across different parallel machines. Our initial results are very good both in terms of performance and scalability. We have been able to obtain a latency of 350 milliseconds with a throughput of 7.5 inputs per second for the entire application using 234 processors on the Intel Paragon deployed at Rome Laboratory. The results have clearly demonstrated the scalability of our approach not only for individual tasks but also for the entire application both in terms of latency and throughput. Many new optimizations are currently being designed and will be incorporated in the future.
Given that the STAP application that is parallelized is one of the few used in the DoD (Rome Laboratory has successfully implemented this STAP algorithm on-board an airborne platform), and is one of the most computationally intensive signal processing algorithm with complex data and communication patterns, our project has demonstrated that techniques developed as part of this project are important and high-performance parallel computers can provide significant performance benefits for such applications.