Omni-Path

Omni-Path
Name	Omni-Path
Developer	Intel Corporation
Type	High-performance computing interconnect
Introduced	2015
Discontinued	2019 (product line shift)
Media	Optical, copper
Predecessor	InfiniBand
Successor	Intel Ethernet adapters

Omni-Path Omni-Path was a high-performance interconnect architecture developed by Intel Corporation for supercomputing centers such as Argonne National Laboratory, Oak Ridge National Laboratory, and academic facilities like Lawrence Berkeley National Laboratory. Designed to compete with InfiniBand and complement Intel Xeon and Intel Xeon Phi processor deployments, it targeted clusters participating in projects funded by agencies including the United States Department of Energy, the National Science Foundation, and collaborative programs with vendors like Hewlett Packard Enterprise, Cray Inc., and Dell EMC. Early demonstrations involved systems at Los Alamos National Laboratory, Sandia National Laboratories, and procurement by institutions such as Riken, CEA, and CNRS.

Overview

Omni-Path defined a fabric and portfolio of adapters, switches, routers, and cables intended for petascale-class systems used by organizations like Argonne Leadership Computing Facility, Oak Ridge Leadership Computing Facility, and the European Centre for Medium-Range Weather Forecasts. It emphasized low latency and high bandwidth for workloads run on processors produced by Intel Corporation and accelerated by devices such as NVIDIA Tesla, AMD Radeon Instinct, and coprocessors like Intel Xeon Phi (Knights Landing). Competitive comparisons were drawn against technologies from Mellanox Technologies, Cisco Systems, and standards bodies such as the InfiniBand Trade Association. Procurement and deployment decisions often involved integrators like Hewlett Packard Enterprise, Cray Inc., Fujitsu, and Lenovo.

Architecture and Components

The architecture incorporated switch ASICs, host channel adapters (HCAs), and cabling choices supporting topologies used by centers like Argonne National Laboratory and corporations like IBM. Fabric elements included leaf-spine and fat-tree configurations adopted in systems built by Hewlett Packard Enterprise, Cray Inc., and Dell EMC. Key silicon and board partners included Intel Corporation design teams and manufacturing partners such as 台积电 (TSMC) and fabs used by Intel Foundry Services; vendors supplied optics from companies like Finisar and II-VI Incorporated. Management tools interfaced with cluster managers and schedulers such as SLURM, PBS Professional, TORQUE, and LSF from IBM. Interoperability testing involved labs and organizations like NERSC and procurement consortia including PRACE.

Performance and Use Cases

Omni-Path targeted latency- and bandwidth-sensitive workloads typical of codes run by research groups at Los Alamos National Laboratory, Lawrence Livermore National Laboratory, and university centers such as Stanford University and Massachusetts Institute of Technology. Use cases included computational fluid dynamics employed by teams collaborating with NASA, large-scale molecular dynamics used by researchers at Argonne National Laboratory and Riken, climate modeling performed by institutions like NOAA and ECMWF, and big-data analytics in projects funded by DARPA and the National Institutes of Health. Benchmarks compared Omni-Path to Mellanox HDR, 10/40/100 Gigabit Ethernet, and proprietary fabrics in codes such as LAMMPS, GROMACS, NAMD, CP2K, and WRF, often evaluated on interconnect-focused suites developed at NERSC and OLCF.

Software Stack and Programming Model

The software stack comprised low-level firmware and drivers developed by Intel Corporation alongside middleware and libraries like MPI implementations (for example, Open MPI and MVAPICH2), RDMA stacks, and tooling integrating with monitoring systems developed by entities such as Ganglia and Prometheus. Programming models employed by users included message-passing paradigms pioneered in projects at Lawrence Berkeley National Laboratory and partitioned global address space concepts used in research from Sandia National Laboratories; MPI tuning and collective algorithms referenced work from EPCC, NERSC, and compiler teams at Intel Corporation. Deployment often required orchestration with software from SUSE Linux Enterprise Server, Red Hat Enterprise Linux, container platforms influenced by Docker, and scheduler plugins contributed by vendor collaborations with HPE and research labs.

History and Development

Omni-Path originated from strategic initiatives within Intel Corporation after acquisitions and competitive positioning relative to Mellanox Technologies, Cisco Systems, and legacy adopters of InfiniBand. Public announcements occurred at venues such as the SC Conference and ISC High Performance. Development engaged partnerships and procurements with national labs including Argonne National Laboratory, Oak Ridge National Laboratory, Lawrence Berkeley National Laboratory, and collaborations with system vendors Cray Inc., Hewlett Packard Enterprise, and Dell EMC. Over time, market dynamics and corporate strategy led to shifts toward Ethernet-based offerings and changing focus among procurement groups such as DOE facilities and European agencies like PRACE and EuroHPC.

Adoption and Industry Impact

Adoption included deployments at national laboratories like Argonne National Laboratory and research centers such as NERSC; collaborations influenced decisions by supercomputing projects funded by DOE and NSF. The product shaped discussions among vendors including Mellanox Technologies, Cisco Systems, Hewlett Packard Enterprise, and integrators like Cray Inc. about interconnect roadmaps, and informed standards conversations in communities around InfiniBand Trade Association and high-performance networking research at institutions such as University of Illinois Urbana–Champaign and University of California, Berkeley. Long-term impact influenced vendor consolidation, procurement strategy at centers like Oak Ridge National Laboratory, and technical evolution toward high-radix switches and Ethernet convergence pursued by companies like Broadcom, Intel Corporation, and Arista Networks.

Category:High-performance computing