Accelerare hardware pentru sisteme de calcul distribuit eficiente energetic / Hardware offloading for energy-efficient distributed computing

This research performed in this thesis is focused on finding practical methods for obtaining energy-efficiency in distributed computing systems. We implemented a distributed computing heterogeneous cluster based on the Hadoop framework with hardware-accelerated GZIP compression. The cluster is composed of ZedBoard development boards, which contain dual core ARM processors and programmable logic.The research included three stages. In the first stage, we compared the performance and the energy-efficiency of the ARM based cluster with an x86 cluster while running standard benchmarks in the Hadoop framework. The ARM based cluster is 5x times more energy efficient than the x86 cluster. In the second stage of our research, we developed the first open source, technology independent GZIP FPGA core using the Verilog Hardware Description Language. The core can reach up to 1.3 Gbps and is up to 18x times faster than an Intel i7 running at a frequency 28x times faster. The third stage of our research focused on integrating the FPGA compression core in the Hadoop ZedBoard cluster. The FPGA GZIP core is 5.6x times faster compared to the ARM CPU which has a clock frequency 6.7x times higher. We obtained a cluster which is almost 2x times more energy efficient when the GZIP hardware compression is used during the Map stage of Hadoop framework.