This article was accepted into the corpus but its outbound wikilinks were never NER-processed — typical at the deepest BFS hop or when the run's entity cap was reached. No expansion funnel to show.
| Myrinet | |
|---|---|
| Name | Myrinet |
| Developed | Myricom |
| Introduced | 1990s |
| Type | high-speed local area network |
| Bandwidth | 1 Gbit/s (initial) |
| Successor | InfiniBand |
| Usage | high-performance computing, clusters |
Myrinet
Myrinet was a high-performance fabric designed for cluster computing and scientific applications, developed by Myricom and deployed in supercomputing centers, university laboratories, and industrial research facilities. It competed with contemporaries in the late 1990s and early 2000s and influenced technologies used by projects at institutions such as Lawrence Livermore National Laboratory, Argonne National Laboratory, Los Alamos National Laboratory, Oak Ridge National Laboratory and commercial vendors like Cisco Systems, Sun Microsystems, Intel Corporation and IBM. The design served workloads from computational fluid dynamics used by researchers at NASA and Duke University to molecular dynamics in groups at Stanford University and Massachusetts Institute of Technology.
Myrinet’s origins trace to research and development at Myricom and collaborations with academic groups and national laboratories including University of California, Berkeley, California Institute of Technology, Princeton University and University of Illinois Urbana-Champaign. Early deployments appeared alongside clusters built with servers from Silicon Graphics, Dell Technologies, HP (Hewlett-Packard), and specialty systems by Cray Research. Funding and interest intersected with initiatives such as the National Science Foundation distributed computing programs, procurement at Los Alamos National Laboratory, and vendor consortia involving Intel Corporation and Microsoft. Competitive and complementary technologies included projects by Larry Smarr’s teams, protocols from Apache Software Foundation researchers, and hardware efforts tied to DARPA and Department of Energy research portfolios. Over time, market shifts toward standards like InfiniBand, emerging Ethernet enhancements championed by Broadcom and acquisitions influenced Myricom’s trajectory as major players such as HP and IBM emphasized alternative fabrics.
The fabric employed a low-latency, high-throughput topology with routing and flow-control mechanisms influenced by research from groups at Carnegie Mellon University, MIT Lincoln Laboratory, and University of Cambridge. Its switched architecture supported cut-through routing and credit-based flow control comparable to work at Stanford University’s networking labs and models from Bell Labs research. Designs emphasized kernel bypass and user-level networking techniques championed by teams at University of California, San Diego and researchers associated with Andrew Yao and colleagues. The architecture accommodated collective communication patterns similar to MPI implementations used at Argonne National Laboratory and Oak Ridge National Laboratory and integrated with parallel file systems developed by teams at University of California, Santa Cruz and Los Alamos National Laboratory.
Myrinet systems used interface cards, switches, and cabling produced by Myricom and compatible vendors such as Emulex, QLogic, Broadcom, and server OEMs including Dell Technologies and Sun Microsystems. Host channel adapters functioned similarly to HCA designs later adopted by InfiniBand Trade Association members and paralleled NIC research at Intel Corporation and Broadcom Corporation. Switch fabrics echoed concepts from Cisco Systems and crossbar ideas from Sequent Computer Systems and Thinking Machines Corporation. Physical interconnects and transceivers met standards used in datacenter equipment from Fujitsu, NEC, and Hitachi.
Software stacks for the fabric integrated with operating systems such as Linux, Solaris, Microsoft Windows NT (in lab ports), and research kernels at University of Washington. MPI implementations from Argonne National Laboratory (MPICH), and libraries from Open MPI contributors were ported to exploit Myricom’s user-level networking APIs. Middleware and resource managers from SLURM Workload Manager, PBS Professional, and Torque (software) supported clusters using the fabric, while storage integration used parallel file systems like Lustre and GPFS developed by IBM. Protocol innovations paralleled academic work at ETH Zurich and EPFL on zero-copy and RDMA primitives.
Myrinet delivered low latency and high bandwidth for distributed applications studied in benchmarks at Lawrence Berkeley National Laboratory and Sandia National Laboratories. Performance claims were validated in comparative studies involving InfiniBand, enhanced Ethernet research from Intel, and clusters at National Center for Supercomputing Applications. Scalability experiments mirrored topologies used in systems at Texas Advanced Computing Center, Pittsburgh Supercomputing Center, and university clusters at Cornell University and Yale University. The fabric’s suitability for tightly-coupled HPC codes was assessed alongside implementations of LAPACK and ScaLAPACK used by teams at Argonne National Laboratory and Oak Ridge National Laboratory.
Deployments occurred in scientific computing centers, enterprise research labs, and commercial clusters built by Sun Microsystems, Cray Inc., Dell Technologies, and integrators working with Myricom. Applications included climate modeling at National Center for Atmospheric Research, computational chemistry at Rice University and University of Illinois, astrophysics simulations at Princeton University, and finance risk models in firms linked with Goldman Sachs research teams. Integration with resource managers from SLURM Workload Manager and visualization pipelines used by Visualization Sciences Group and academic visualization labs at University of Utah showcased the fabric’s versatility.
Myrinet’s influence appears in later high-performance fabrics such as InfiniBand and enhanced Ethernet deployments driven by groups at Intel Corporation, Broadcom Corporation, and standards bodies including IEEE. Its concepts informed user-level networking, RDMA, and low-latency switching work at Microsoft Research, Google Research, Facebook (Meta Platforms, Inc.) data-center networking teams, and academic groups at ETH Zurich and University of Cambridge. Vendors including Mellanox Technologies (now part of NVIDIA), QLogic, and Emulex built on similar principles to serve HPC and cloud markets, while supercomputing centers such as Oak Ridge National Laboratory and Argonne National Laboratory migrated to newer fabrics for exascale initiatives sponsored by Department of Energy programs and collaborations with Cray Research/Hewlett Packard Enterprise and IBM.