Network Synthesis for Database Processing Units

We explore on-chip network topologies for the Q100, an analytic query accelerator for relational databases. In such data-centric accelerators, interconnects play a critical role by moving large volumes of data. In this paper we show that various interconnect topologies can trade a factor of 2.5x in performance for 3.3x area. Moreover, standard topologies (e.g., ring or mesh) are not optimal. Significant prior work on network topology specialization augments generic topologies with additional dedicated links. In this paper, we present a network specialization algorithm that builds a specialized network first then introduces a generic network as a fallback. We find our algorithm produces networks that are 1.24x slower than the highest-performance generic topology considered (a fat tree), and 18% smaller than the least expensive (a double ring). Moreover, our method produces topologies that outperform those produced by others by 1.21x while being 25% smaller.