Programming Constructs for Exascale Computing in Support of Space Situational Awareness

摘要

Increasing data burdens associated with image and signal processing in support of space situational awareness implies much-needed growth of computational throughput beyond petascale (10~(15) FLOP/s) to exascale regimes (10~(18) FLOP/s, 10~(18) bytes of memory, 10~(18) disks and Input/Output (I/O) channels, etc.) In addition to growth in applications data burden and diversity, the breadth and diversity of high performance computing architectures and their various organizations have confounded the development of a single, unifying, practicable model of parallel computation. Therefore, models for parallel exascale processing have leveraged architectural and structural idiosyncrasies, yielding potential misapplications. In response to this challenge, we have developed a concise, efficient computational paradigm and software called Program Compliant Exascale Mapping (PCEM) to facilitate efficient optimal or near-optimal mapping of annotated application codes to parallel exascale processors. Our theory, algorithms, software, and experimental results support annotation-based parallelization of application codes for envisioned exascale architectures, based on Image Algebra (IA) [RitOl]. Because of the rigor, completeness, conciseness, and layered design of image algebra notation, application-to-architecture mapping is feasible and scalable at exascales., In particular, parallel operations and program partitions are categorized in terms of six types of parallel operations,where each type is mapped to heterogeneous exascale processors via simple rules in the PCEM annotation language. In this paper, we overview opportunities and challenges of exascale computing for image and signal processing in support of radar imaging in space situational awareness applications. We discuss software interfaces and several demonstration applications, with performance analysis and results in terms of execution time as well as memory access latencies and energy consumption for bus-connected and/or networked architectures. The feasibility of the PCEM paradigm is demonstrated by addressing four principal challenges::(1) architectural/structural diversity,parallelism, and locality, (2) masking of I/O and memory latencies, (3) scalability, and (4) efficient representation/expression of parallel applications. Examples will demonstrate how PCEM helps solve these challenges efficiently on real-world computing systems;