Slingshot (interconnect)

Slingshot (interconnect). It is a high-performance, high-radix Ethernet-based interconnect network designed for exascale computing and large-scale high-performance computing (HPC) systems. Developed by Hewlett Packard Enterprise (HPE) through its Cray division, Slingshot serves as the foundational networking technology for several of the world's most powerful supercomputers. The interconnect employs a novel Dragonfly+ topology and specialized features to efficiently handle the demanding, mixed workloads of modern scientific computing, including traditional HPC simulations, big data analytics, and artificial intelligence workloads.

Overview

Slingshot represents a significant evolution from previous Cray interconnects like Aries and Gemini, moving to a standards-based Ethernet architecture while retaining and enhancing performance characteristics critical for exascale computing. It was selected as the interconnect for the United States Department of Energy's (DOE) first exascale systems, including the Frontier system at Oak Ridge National Laboratory and the El Capitan system at Lawrence Livermore National Laboratory. The design philosophy centers on providing extremely low latency, high bandwidth, and advanced Quality of Service (QoS) features to prevent network congestion from impacting application performance in massive, multi-tenant environments shared by thousands of simultaneous jobs.

Architecture

The architectural core of Slingshot is its adaptive routing and congestion control within a high-radix Dragonfly+ network. This topology groups compute nodes into router-based cabinets, with optical links connecting groups into a global network, minimizing the number of expensive long-distance hops. Key innovations include the use of 64b/66b encoding and a link-level retransmission protocol for reliability. The interconnect implements sophisticated per-packet Quality of Service mechanisms and explicit congestion notification, allowing it to isolate traffic and prevent network congestion caused by a single application from degrading the performance of others running on the same system, a critical capability for cloud computing-style environments on supercomputers.

Performance and capabilities

Each Slingshot network interface controller (NIC) port supports a data rate of 200 gigabits per second (Gb/s), with each compute node typically utilizing multiple ports for aggregate bandwidth. The interconnect demonstrates exceptionally low switch latency, measured in nanoseconds, and provides a sustained bisection bandwidth scaling into hundreds of petabits per second for full-system configurations. Its advanced Quality of Service features support up to 32,000 virtual channels, enabling fine-grained traffic isolation for diverse workloads from computational fluid dynamics to machine learning training. This performance has been instrumental in systems like Frontier achieving first place on the TOP500 and HPL-AI benchmarks, validating its capability to support real-world exascale scientific computing applications.

Deployment and usage

Slingshot is deployed in several of the world's leading supercomputers, most notably the exascale systems built for the United States Department of Energy. The Frontier system at Oak Ridge National Laboratory, which became operational in 2022, was the first to publicly demonstrate exascale performance using Slingshot. It is also the interconnect for the Aurora system at Argonne National Laboratory and the forthcoming El Capitan system at Lawrence Livermore National Laboratory. Beyond these flagship systems, Slingshot technology is integrated into HPE's commercial HPC offerings, such as the HPE Cray EX supercomputer, making it available for research institutions, government agencies, and commercial entities worldwide for demanding workloads in fields like climate modeling, nuclear weapons simulation, and pharmaceutical research.

Development and history

The development of Slingshot was led by Hewlett Packard Enterprise following its acquisition of Cray Inc. in 2019, building upon decades of Cray's expertise in designing proprietary interconnects for vector processor and massively parallel systems. The project was driven by the specific requirements of the United States Department of Energy's Exascale Computing Project (ECP), which sought a new networking paradigm capable of handling exascale workloads. Key design work and validation were conducted in partnership with DOE national laboratories, including Lawrence Berkeley National Laboratory. The first production deployment was in the Frontier system, with its public unveiling in 2022 marking a milestone in high-performance computing history and establishing Slingshot as a foundational technology for the exascale era.

Category:Computer network hardware Category:Supercomputer interconnects Category:Hewlett Packard Enterprise