Heterogeneous Servers based on Programmable Cores and Dataflow Engines
Published in Workshop on Energy-Efficient Servers for Cloud and Edge Computing 2017, 2017
This paper addresses the challenges of increasing data volumes and stringent power budgets in modern data centers by proposing heterogeneous server architectures that integrate programmable cores and dataflow engines. The authors present two architectural approaches: one based on reconfigurable tiny cores and another utilizing dataflow engines, both aimed at enhancing energy efficiency and performance in cloud and edge computing environments.
Key contributions include:
Nanocore Architecture: Development of a micro-server architecture incorporating reconfigurable tiny cores optimized for FPGA platforms, achieving up to 40% better energy efficiency compared to standard Xeon servers.
Dataflow Engine Integration: Exploration of programmable dataflow engines, demonstrating speedups of up to 374x for various workloads, highlighting their potential for high-throughput data processing.
NanoWire Communication Protocol: Introduction of a low-latency communication protocol, NanoWire, enabling efficient data transfer between host processors and accelerators, facilitating scalable and flexible system architectures.
Recommended citation: Wu, Y., Gillan, C., Minhas, U., Barbhuiya, S., Novakovic, A., Tovletoglou, K., Tzenakis, G., Vandierendonck, H., Karakonstantis, G., & Nikolopoulos, D. (2017). Heterogeneous Servers based on Programmable Cores and Dataflow Engines. In *Workshop on Energy-Efficient Servers for Cloud and Edge Computing 2017*. Queen’s University Belfast. https://pureadmin.qub.ac.uk/ws/files/123397592/HiPEAC_ESCEC_2016.pdf
Download Paper | Download Slides | Download Bibtex