Cray Aries

Cray Aries
Name	Aries
Manufacturer	Cray Inc.
Introduced	2012
Type	Supercomputer interconnect / network
Architecture	Dragonfly topology, PCIe, 3D torus influence
Topology	Dragonfly
Speed	40 Gbps per port (varied implementations)
Predecessors	Gemini
Successors	Slingshot

Cray Aries

Cray Aries is a high-performance interconnect developed by Cray Inc. that provided low-latency, high-bandwidth networking for Cray supercomputers and clusters. It combined custom routing, adaptive flow control, and a Dragonfly-inspired topology to support large-scale deployments for computational science, weather forecasting, and national laboratory workloads. Aries was integrated into several systems and influenced subsequent designs in high-performance computing and data-center fabrics.

Overview

Aries served as the packet-switched, RDMA-capable fabric for Cray systems, addressing needs of projects such as Oak Ridge National Laboratory, Los Alamos National Laboratory, Argonne National Laboratory, National Center for Atmospheric Research, and commercial partners like Sandia National Laboratories collaborators. Designed to interoperate with MPI implementations used by researchers at Lawrence Livermore National Laboratory, European Centre for Medium-Range Weather Forecasts, and universities such as Massachusetts Institute of Technology and Stanford University, Aries targeted workflows in climate modeling, computational chemistry, and seismic analysis. It competed in the market alongside interconnects from InfiniBand Trade Association vendors, Intel networking efforts, and projects involving NVIDIA and Broadcom technologies.

Architecture and Design

The Aries design integrated custom ASICs, routers, and host adapters to provide a flat, low-diameter network supporting RDMA semantics used by MPI stacks like Open MPI and MPICH. Hardware components interfaced to compute nodes via PCI Express and supported system management from tooling similar to that used by Hadoop clusters and cluster management frameworks at sites like Lawrence Berkeley National Laboratory. Aries incorporated credit-based flow control and congestion management strategies that paralleled research by groups at University of California, Santa Cruz and University of Illinois Urbana-Champaign. The interconnect's silicon drew on circuit-design practices from collaborators including Cadence Design Systems and Synopsys IP flows.

Routing and Interconnect Technology

Aries implemented a Dragonfly-inspired topology with global and local group links, enabling low hop-count paths between endpoints in large systems deployed at institutions like Princeton University and Yale University. The routing logic supported adaptive routing algorithms similar to those studied by researchers at Massachusetts Institute of Technology and Cornell University to mitigate hot-spots encountered in scientific runs from applications such as GROMACS, LAMMPS, and NAMD. Aries routers provided virtual channels and Quality of Service features that allowed batch-schedulers like SLURM and Torque to manage network-aware job placement. The technology addressed congestion scenarios discussed in conferences hosted by ACM and IEEE.

Performance and Scalability

Aries delivered low latency and high bandwidth for large MPI collectives used in codes developed at Princeton Plasma Physics Laboratory and Fermi National Accelerator Laboratory. Benchmarks run by teams at Argonne National Laboratory and Oak Ridge National Laboratory demonstrated scalability across thousands of nodes for applications including WRF, CESM, and NWChem. The interconnect's aggregate bisection bandwidth and adaptive routing enabled sustained performance under communication-heavy kernels used by researchers from Caltech and Georgia Institute of Technology. Aries' deployment in large systems informed procurement and architecture discussions at institutions such as National Energy Research Scientific Computing Center.

Deployment and Integration

Cray integrated Aries into its XC-series systems and offered it as an option for cluster deployments by national laboratories and universities including Brookhaven National Laboratory and IBM-collaborative centers. System software stacks incorporated Aries support into resource managers like Grid Engine and monitoring suites used at European Organization for Nuclear Research computing facilities. Integration work involved collaborations with compiler and runtime projects at University of California, Berkeley and performance-analysis tools from teams at Los Alamos National Laboratory and Sandia National Laboratories.

History and Development

Aries succeeded Cray’s Gemini interconnect and represented an evolution influenced by public research on low-diameter topologies from groups at University of Washington and University of Texas at Austin. Development was driven by Cray engineering groups and partnerships with government procurement programs at Department of Energy laboratories and international research centers such as CINECA and PRACE. The interconnect entered production in the early 2010s and was featured in procurement announcements alongside systems from vendors such as Hewlett Packard Enterprise and Dell EMC.

Legacy and Impact

Aries left a legacy in informing the design of successor fabrics and commercial network products, influencing later efforts from Cray Research alumni and acquisitions involving HPE and networking roadmaps at NVIDIA and AMD. Lessons from Aries’ routing, flow control, and system integration contributed to standards discussions at the OpenFabrics Alliance and academic publications presented at SC Conference and International Supercomputing Conference. Its deployment supported major scientific milestones at laboratories like Oak Ridge National Laboratory and helped shape large-scale HPC procurement and architecture choices through the 2010s and beyond.

Category:Supercomputer interconnects