ACADEMIA
NASA, Livermore Lab Confirm Obsidian’s BGFC First Order Functionality
BGFC Will Help Supercomputing Get Bigger – Much Bigger!
In less than a week, attendees to the annual Supercomputing Conference in Seattle, Washington will witness first hand how BGFC will push the limits of InfiniBand architectures into the next generation of scalability.
Future ExaScale-class supercomputer architectures demand segmented interconnect networks for performance, scalability and fault isolation. All subnet managers available today only support single-subnet InfiniBand networks, yet very large single subnets are already presenting scaling challenges in the largest systems.
Lawrence Livermore National Laboratory (LLNL) and NASA Ames Research Center have collaborated with Obsidian Research in the specification and verification of a next-generation InfiniBand subnet manager – BGFC. A green field re-imagining of the InfiniBand subnet manager, BGFC features ExaScale-class characteristics such as mathematically verifiable graph-theory based routing algorithms, direct support for InfiniBand routers and the ability to cluster the subnet manager functions locklessly for high-availability at extreme-scale or across geo-distributed multi-subnet fabrics.
LLNL is a National Nuclear Security Administration (NNSA) science and technology laboratory that deploys some of the world’s most powerful supercomputers to ensure the safety, security and reliability of the nation’s nuclear deterrent without underground testing.{dmMaps strt="1320866429" DO NOT EDIT/REMOVE, HIDDEN ON THE FRONTEND ;) }
In this technology preview of the GPL-licensed BGFC, a geo-distributed two-subnet InfiniBand fabric is established spanning LLNL and NASA Ames using Obsidian's Longbow E100 range-extending routers. Various forms of InfiniBand traffic, including IPoIB, are passed at wire speed with AES encryption between LLNL's Hyperion cluster and NASA's Pleiades nodes (this demonstration is notable since it includes a complex topology, combining Clos and Hypercube subnets into a single fabric).
Live demonstrations will take place in the NNSA’s Advanced Simulation and Computing (ASC) program booth #803 at 5 pm, 2 pm and 12 noon on November 15th,16th and 17th respectively – senior network administrators will be on hand during demonstration times to field questions, offer their thoughts and vision for the future of scalable HPC cluster interconnects. A brief paper authored by Obsidian’s architects will be made available onsite to help potential users of BGFC prepare for its release to the open community, expected in mid-2012.
In the meantime, a second phase of field testing is scheduled to begin immediately following SC11 to expose BGFC's architecture to a wider selection of supercomputer and storage fabric topologies and environments.