Dynamic Data Driven SAR Reconstruction on Hybrid Multicore systems

Abstract

The reconstruction of nxn-pixel Synthetic Aperture Radar imagery using a Backprojection algorithm is compute intensive and incurs O(n2 · m) cost, where m is the number of pulses. As part of this research, we develop dynamic data driven multiresolution algorithms that speed up SAR backprojection on GPUs, hybrid multicore and many-core processors. Further, we performed experiments to observe improvements on a variety of architectures.The challenges in improving performance of this spatially variant reconstruction process on any architecture is load balancing, which circumvents asymmetric work assignment. On GPUs, fine tuned algorithms were developed as part of our research for improving execution time. Further, communication between processors was overlapped with computation to reduce overall execution time.We also developed parallel algorithms and software for constructing multi-resolution SAR images on hybrid multicore processors (HMPs). In particular, several load balancing algorithms were developed for optimizing performance and energy consumption on HMPs. We also developed a systematic approach for deriving the performance-energy trade-offs on HMPs while exploiting dynamic voltage and frequency scaling (DVFS) features of CPU cores and GPUs. This approach helps the user to select the right system configuration, that is, the number of processing elements of each type (cores/GPUs/etc.) and the respective clock frequencies, depending on whether performance or energy optimization is critical to the user.We evaluated performance and energy consumption of our algorithms on an Intel Knights Landing (KNL) processor as a representative of a many-core architecture. We also compared performance and energy consumption of KNL, Ivy Bridge and Tesla K40m.