High-level synthesizable dataflow MapReduce accelerator for FPGA-coupled data centers

Abstract

Manipulating big-data entries of emerging server workloads requires a design paradigm shift towards more aggressive system-level architecture solutions. From software perspective, the MapReduce framework is a prominent parallel data processing tool as the volume of data to analyze grows rapidly. FPGAs can be used to accelerate the processing of data and reduce significantly the power consumption. However, FPGAs have not been deployed in data centers due to the high programming complexity of hardware. In this paper we present HLSMapReduceFlow, i.e. a novel reconfigurable MapReduce accelerator that can be scaled-up to data centers and it can speedup the processing of Map computation kernels, while promising minimum energy footprint and high programming efficiency due to the use of HLS. We propose the complete decoupling of MapReduce's tasks data-paths to distinct buses, accessed from individual processing engines. Such a dataflow approach implies a holistic C/C++ to RTL domain-level MapReduce transition. In this work, we further extent HLS tools, with systematic source-to-source code annotation of HLS optimization directives, by adding as a state-of-art system-level implementation toolflow. The proposed architecture is implemented, mapped and evaluated to a Virtex-7 FPGA and shows that the proposed scheme can achieve up to 4.3× overall throughput improvement in MapReduce applications, while offering two orders of magnitude power/energy improvements compared to a high-end multi-core processor.