针对SAR成像核心算法可配置软加固技术研究

发布时间：2018-08-11 11:03

【摘要】：硬件容错加固技术,具有能耗、面积及复杂性线路等方面的开销,因此对于星载应用,硬件容错代价高昂。软件容错技术应运而生,其良好的灵活性、可配置性及较高的错误覆盖率,很好满足了星载软件的发展需要。软件容错由于无需修改硬件,而被用于可靠性加固。随着软件容错技术地不断发展,出现了大量性能良好的容错算法——无论在控制流或数据流方面,都形成了许多思路方法巧妙且检错能力强的算法。然而,算法也产生了运算开销与加固强度的折衷选择问题。因此,针对特定算法特点ABFT(Algorithm Based Fault Tolerance),且作可配置软件加固CBFT(Configurable Based Fault Tolerance)技术,受到了特别青睐。本文在借鉴软加固经典算法的同时,针对RDA算法特点,提出了一种针对运算特征、简单而又满足运算性能需求的多层次软加固技术。对此,我们作了几个方面的讨论:1.针对RDA成像核心部件——FFT运算,文章提出基于数据加权且校验和的容错算法WCFT。文章并分析了算法的错误检测能力、性能开销,算法设计的特点及优点。2.文章提出针对RDA程序特点的ABFT算法,整合校验和方法及WCFT算法的特点,给出一种有效的错误检测恢复机制。针对运算精度要求,权衡性能开销,提出双精度运算及三角函数置表方法;针对FFT中错误高发部分——倒序运算,提出倒序置表方法;提出针对算法增量运算的错误检测方法,及针对全局变量的冗余检测方法。3.针对多处理器条件下的并行运算,文章分析了冗余多线程执行问题及内存访问竞争问题。文章并对增量因子错误问题作了分析,针对这些问题提出相关技术改进方法。4.采用同时输入多行/列数据的FFT运算——SIMD(Single Intruction Multiple Data)[53]设计,使性能得到近一倍提升,并基于此提出针对性的错误检测恢复机制。同时该方法,节省了四次矩阵转置操作,减少了内存数据迁移及数据校验开销。5.根据不同的应用需求级别,提出针对SAR程序的CBFT算法:作了两种级别的加固。文章并使用冗余算法作了错误检测恢复实现。
[Abstract]:Hardware fault-tolerant reinforcement technology has the cost of energy consumption, area and complex circuits, so hardware fault tolerance is very expensive for space-borne applications. Software fault-tolerant technology emerges as the times require, it has good flexibility, configurable and high error coverage, which can meet the development needs of space-borne software. Software fault tolerance is used for reliability reinforcement because there is no need to modify hardware. With the continuous development of software fault-tolerant technology, a large number of fault-tolerant algorithms with good performance have emerged. In both control flow and data flow, many algorithms with clever thinking methods and strong error detection ability have been formed. However, the algorithm also produces a trade-off between computational overhead and reinforcement strength. Therefore, CBFT (Configurable Based Fault Tolerance) technology based on ABFT (Algorithm Based Fault Tolerance), and configurable software is very popular. According to the characteristics of RDA algorithm, this paper presents a multi-level soft reinforcement technique which is simple and meets the requirement of operation performance. In this regard, we made several aspects of the discussion: 1. In this paper, a fault tolerant algorithm based on data weighted and checksum is proposed for the RDA imaging core component. This paper also analyzes the error detection ability, performance overhead, the characteristics and advantages of the algorithm design. This paper presents an effective error detection and recovery mechanism based on the characteristics of ABFT algorithm, integrated checksum method and WCFT algorithm, which is based on the characteristics of RDA program. According to the requirement of operation precision and weighing the performance overhead, this paper puts forward the double precision operation and trigonometric function table setting method; for the high error occurrence part-inversion operation in FFT, the inverted table method is put forward; and the error detection method for incremental algorithm operation is put forward. And the redundancy detection method for global variables. 3. In this paper, the problem of redundant multithread execution and memory access competition is analyzed for parallel operations under multiprocessor conditions. In this paper, the error problem of increment factor is analyzed, and the relative technical improvement method is put forward. 4. SIMD (Single Intruction Multiple Data) [53] is designed by using FFT operation which inputs multi-row / column data at the same time, and the performance is nearly doubled. Based on this, a targeted error detection and recovery mechanism is proposed. At the same time, this method saves the fourth order matrix transpose operation, reduces the memory data migration and data check cost. According to different application requirement levels, a CBFT algorithm for SAR programs is proposed: two levels of reinforcement are made. In this paper, the redundant algorithm is used to implement the error detection and recovery.
【学位授予单位】：国防科学技术大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TN957.52

【参考文献】