Frederick Ryan Johnson, Ippokratis Pandis
Software has always ruled database engines, and commodity processors riding Moore's Law doomed database machines of the 1980s from the start. However, today's hardware landscape is very different, and moving in directions that make database machines increasingly attractive. Stagnant clock speeds, looming dark silicon, availability of reconfigurable hardware, and the economic clout of cloud providers all align to make custom database hardware economically viable or even necessary. Dataflow workloads (business intelligence and streaming) already benefit from emerging hardware support. In this paper, we argue that control flow workloads with their corresponding latencies are another feasible target for hardware support. To make our point, we outline a transaction processing architecture that offloads much of its functionality to reconfigurable hardware. We predict a convergence to fully "bionic" database engines that implement nearly all key functionality directly in hardware and relegate software to a largely managerial role.
2013
