FPGA-accelerated Heterogeneous Hyperscale Server Architecture for Next-Generation Compute Clusters
René Griessl, Meysam Peykanu, Jens Hagemeyer, Mario Porrmann, Stefan Krupop, Micha vor dem Berge, Lars Kosmann, Patrick Knocke, Michał Kierzynka, Ariel Oleksiak · PUB – Publications at Bielefeld University (Bielefeld University) · 2015
Hyperscale servers tend to increase the resource efficiency in data centres by utilizing specialized and highly modular platforms.Instead of providing extreme high-performance nodes, scale-out is the main paradigm to increase the overall performance, i.e., adding additional energy-efficient compute nodes via high-bandwidth low-latency networks.In addition to this, heterogeneous architectures that can be tailored towards the specific needs of a particular application by integrating, e.g., FPGAs and GPUs, are a promising approach towards resource-efficient high-performance computing.In this paper, we present a novel highly-scalable, heterogeneous server architecture that seamlessly integrates arbitrary combinations of microservers based on general purpose CPUs, low power mobile CPUs, FPGAs and GPUs.Mobile CPUs based on the latest ARM Cortex-A15 devices with integrated GPUs are combined with FPGA-based reconfigurable SoCs, which can be used for application-specific hardware acceleration.A flexible multi-level interconnect enables communication between the microservers based on standard protocols like 10 Gigabit Ethernet.Additionally, serial high-speed links between the FPGA-based microservers are used as communication accelerators for highbandwidth, low-latency data transmission.Control and finegrained monitoring of all relevant system parameters is enabled by a dedicated monitoring network.The performance and energy efficiency of the platform is evaluated based on a set of synthetic benchmarks as well as a real-world sequence alignment application, which shows an increase in energy efficiency compared to CPU and GPU implementations.