Leveraging Reconfigurable Computing in Distributed Real-time Computation Systems

The community of Big Data processing typically performs real-time computations on data streams with distributed systems such as the Apache Storm. Such systems offer substantial parallelism; however, the communication overhead among nodes for the distribution of the workload places an upper limit to the exploitable parallelism. The contribution of the present work is the integration of a reconfigurable platform with the Apache Storm, which is the main platform of the Big Data streaming processing community. By exploiting the internal bandwidth of FPGAs we show that the computational limits for stream processing are significantly increased vs. conventional distributed processing without compromising on the platform of choice or its seamless operation in a dynamic pipeline. The integration of a Maxeler MPC-C Series platform with the Apache Storm, as presented in detail, yields on the Hayashi-Yoshida correlation algorithm an impressive tenfold increase in real-time streaming input capacity, which corresponds to a hundred-fold computational load. Our methodology is sufficiently general to apply to any class of distributed systems or reconfigurable computers, and this work presents quantitative results of the expected I/O performance, depending on the means of network connection. CCS Concepts •Computer systems organization ➝ Distributed architectures ➝Cloud computing •Computer systems organization➝ Heterogeneous (hybrid) systems, Reconfigurable computing, Data flow architecture