Investigation of Reconfigurable Pipelined Architecture for On-the-fly Processing of Big Data Streams Open Access
Downloadable ContentDownload PDF
Real-time processing is one aspect of the Big Data situation, and it requires an unconventional approach to solving the recent problems that may appear at both software and hardware levels. The continuous increase of data velocity has placed a tremendous pressure on the existing systems, and the current volumes of Big Data limit the efforts of storing everything at a preprocessing stage. The conventional processing infrastructures have been thoroughly challenged by the huge demand for the stream-based applications. The industry responded by introducing traditional stream-processing engines along with emerged technologies. The ongoing paradigm embraces the parallel computing as the most reasonable proposition. Pipelining and Parallelism have been intensively studied in recent years. Nevertheless, the challenges of Big Data call for a new approach to solving both software and hardware problems, and the continuous increase of data velocity requires a new form of software clustering backed by a special hardware organization. Hence, on-the-fly processing is necessary more than before to support the collective efforts towards Big Data advancements.In the context of data-intensive applications, pipelining arises as a typical approach. The recent improvements in Field-Programmable Gate Arrays (FPGAs) opened a new door for application-specific architectures by increasing the usability of FPGA hardware design. This dissertation presents a reconfigurable pipelined architecture which is mainly aimed at Big Data preprocessing. It adopts Symmetric multiprocessing (SMP) along with crossbar switch and forced-interrupt that allows processing of variable-lengths of code.The primary goal of this promising architecture is to process efficiently Big Data streams on-the-fly while it can process sequential programs on a parallel-pipelined model. The system overpasses internal memory constraints of Non-uniform memory access (NUMA) by applying forced-interrupts and crossbar switching. It reduces complexity, data dependency, high-latency, and cost overhead of parallel computing.Undoubtedly, the hardware architecture plays a significant role in improving the efficiency of streaming systems. The variance of hardware performance on different hardware architectures is quite interesting. The previous literature confirms that the CPUs, the GPUs, and the FPGAs are performing differently in specific applications. Based on earlier work, the GPUs outperformed the other platforms in terms of execution time. CPUs outperformed in overall execution combined with transfer time. FPGAs outperformed for fixed algorithms using streaming . Accordingly, this dissertation presents a reconfigurable pipelined architecture that is especially designed to improve the performance of real-time processing, and it examines the performance of three streaming applications on three different hardware platforms; that is followed by the experimental results. The analysis confirms a noticeable improvement by pipelining on the reconfigurable FPGA design.