HPC Problems On SAR Imaging

SAR imaging plays an important role in an SAR processing system and directly brings a greate impact on system throughput and rapid response capability. Compared with other HPC problems, SAR imaging, which presents its own characteristics, must deal a large volume of calculation and data.With an example for Chirp Scaling algorithm, this paper focuses on typical operations in SAR imaging and analyzes the elements of a HPC system, especially of a NUMA system, which greatly impact SAR imaging performance. The results indicate that the serial and parallel optimization for FFT, phase compensation and matrix transpose has the most significance for SAR imaging performance.For serial optimization, theoretic analysis and benchmarking are carried on to improve the floating point performace and to reuse the cache more effectively. Also, the relactions between the best block size for matrix transpose and the size and associativity of the processor's cache is formulized.For parallel optimization, several programming models available on a NUMA system, such as lightweight processes (sproc), POSIX threads, OpenMP and MPI, are compared, and their speedup and coding complexity are analyzed. Among them, OpenMP, which brings competitive performance as lightweight processes, is considered as a better choice than POSIX threads and MPI.