Arista Selected for ECMWF Critical and Extensible Network Infrastructure

• Application Infrastructure & Middleware - Data Exchange & Integration
• Infrastructure as a Service (IaaS) - Cloud Computing
• Infrastructure as a Service (IaaS) - Cloud Storage Services

• Software Design & Engineering Services
• System Integration

The European Centre for Medium-Range Weather Forecasts (ECMWF) needed to ensure its network layer was not a bottleneck for advanced research and weather prediction on upgraded supercomputer clusters. The supercomputer generates a huge amount of data which needs to be offloaded to storage and post-processing as quickly as possible. The network cannot be a bottleneck for this process, and it must cope with predicted upgrades or new workloads. In 2010, with the Centre’s existing IBM-based supercomputers due for upgrade, the team began looking at a possible replacement for its current Force 10 based network.

The European Centre for Medium-Range Weather Forecasts (ECMWF) is an independent intergovernmental organization supported by 34 member states and a world leader in global medium-range numerical weather prediction. Established in 1975, ECMWF employs around 300 staff from more than 30 countries. The Centre is both a research institute and a 24/7 operational service, producing and disseminating numerical weather predictions and data to the national meteorological services in its Member and Co-operating States. ECMWF’s core mission is to produce numerical weather forecasts and monitor the Earth-system, which includes scientific and technical research to improve forecast skill and maintain an archive of meteorological data. Headquartered in Reading, UK, ECMWF provides advanced training to scientific staff in Member and Co-operating States and assists the World Meteorological Organization with its programs. The Centre also offers a catalogue of forecast data that can be purchased by commercial customers worldwide.

The selection process for a new network solution was handled as an open tender, and saw bids from all the main players in the networking space. Arista scored highly, offering some compelling additional benefits. One of these was Arista EOS - a multi-process architecture that separates networking state from the processing itself. This enables fault recovery and incremental software updates on a fine-grain process basis without affecting the state of the system, as well as security patches behind the scenes. Since the selection of Arista in 2010, ECMWF has transitioned through three major supercomputer cluster upgrades, in 2011, 2014, and most recently in 2016. The 2011 installation was almost a like-for-like change, but in 2012 ECMWF deployed a custom EOS extension that unlocked the full potential of the Arista switches and powerful operating systems design. The extensibility features allowed ECMWF to provide transparent failover to supercomputer network nodes without compromising on any design, latency, or operational processes.

• Arista Multi-Chassis Link Aggregation, with Virtual ARP, enables the highest levels of resiliency for 24/7 operation.
• Flexible Arista EOS allows custom on-switch scripting to support supercomputer network node failover.
• Multi-process state sharing architecture allows incremental software updates without affecting the state of the system for improved uptime.

• Combined peak performance of 15.7 petaflops across 260,000 processing cores in two clusters.
• Two Arista 7504-E switches interconnected via 40-Gigabit Ethernet based aggregated links.
• 16 supercomputer network nodes each connected via dual 40-Gigabit Ethernet and all other systems via dual 10-Gigabit Ethernet.