Heterogeneous Servers based on Programmable Cores and Dataflow Engines

The continuous growth of the Internet, and the data volumes associated to it, put severe pressure on the computational, storage and networking resources available in to- day's data-centers. The stringent power budgets and the difficult expansion of current Internet communication infrastructures necessitates the design of new server architectures, programming models and run-time systems. In this paper, we present promising options in the design of energy-efficient servers based on programmable accelerators which can help push the limits on the computation of future data centers. We show that one of the proposed micro-server prototypes based on custom tiny cores can achieve 40% better energy- efficiency than a standard Xeon Server, while data ow engines can help speedup execution by up to 374x for various workloads.