Infiniband and OpenFabrics Alliance Updates
August 16, 2016, Infiniband Trade Association (IBTA) and OpenFabrics Alliance (OFA), San Francisco—William Lee from the IBTA provided an update on the organizations' activities. The IBTA is an organization that develops and promotes standards for interconnect fabrics based on remote direct memory access (RDMA), and supports compliance and compatibility tests. The Open Fabrics Alliance supports and promotes open fabrics software for RDMA for high-performance applications.
The IBTA develops architectural specifications for RDMA, including software architectures and the interconnect fabrics, which are now defined for up to 100Gb/s per port. In addition, the organization is working on RDMA over Converged Ethernet (RoCE). The specifications, technical resources and references are available at RoCEinitiative.org. IBTA is working on a product guide for future release.
Data centers are moving to more parallel processing that requires communications, low latency, and low CPU overhead to achieve their goals of higher throughput. These hyper-converged systems need flexible fabrics to manage key communications. The organization believes that RDMA is essential for most high-performance computing tasks due to the high performance central I/O, which supports scalable storage to meet growing demands, bypass the OS for most data transfers, and improve compute efficiency to reduce power, cooling, and space requirements.
RDMA offers low latency and CPU overhead for efficient data transfer and is OS and protocol neutral. As evidence for their performance claims, they list a 41 percent share of the top 500 clusters, and 70 percent of the high-performance compute clusters. Their users report average compute efficiency of 85 percent with one user showing a use rate of 99.8 percent compared to other technologies.
The IBTA has published a roadmap that takes the fabrics from the existing EDR at 25 Gb/s for a single lane to HDR at 50 Gb/s. Bandwidth is also scalable by adding lanes in multiples of 4.Other projects include developments for more virtualization support, and some vendors are already shipping Infiniband routers.
The OFA is working to deliver open source software to support RDMA in high-performance applications such as enterprise data centers, cloud, and hyper-converged systems. The OFA supports Linux and Windows Server operating systems. This group is establishing itself as a bridge to support both users and developers. The first OFA workshop held in April covered the following seven tracks: communications middleware, distributed applications, emerging technologies, management monitoring and configuration, network APIs and software, network deployment, and networking technologies. Presentations are available at https://openfabrics.org/index.php/2016-ofa-workshop-presentations.html .
In addition to the various sessions, the OFA appointed Susan Coulter from Los Alamos National Lab as the alliance chair and Paul Grun from Cray as the vice chair. This demonstrates their efforts to include users and developers in this important arena in high-performance computing.