iWARP RDMA Ethernet Clustering Enables Large Scale Computing

Chelsio Communications today announced it has completed a major contract with Purdue University to deliver more than 1000 of its Unified Wire 10Gb Ethernet adapters to the Information Technology at Purdue (ITaP) department for the build-out of its high performance computing (HPC) infrastructure, including its large scale “Coates Cluster.”

Tapped as the world’s largest 10Gb Ethernet cluster, the Coates Cluster is expected to be fully operational by the end of July.  ITaP identified numerous requirements for the Coates Cluster to be an effective and innovative computing resource for the entire University:

• only two cables connecting to each server (one power cable, and one network cable)
• standards-compliant Ethernet RDMA clustering performance comparable to Infiniband
• enable fabric convergence of storage, networking, clustering, management and boot traffic on one wire
• support Linux and Windows operating systems
• be cost competitive
•  have independent benchmark data and referenceable installations available for various aspects of the converged network with special emphasis on stability, flexibility, and scalability
• homogeneously connect to other installations at Purdue and other universities
• support various legacy and current application types
• have a roadmap to a motherboard solution to leverage all the software investment that will go into this cluster by various departments

After evaluating all the available technologies and host adapter solutions, only Ethernet and Chelsio’s Unified Wire solution were able to fulfill the stringent requirements.

By separating the purchases of individual cluster components, Purdue was able to select best-of-breed technology to better suit specific research needs.  Funds from a variety of scientific disciplines were pooled together to form a single system, allowing the university to achieve economies of scale during the purchase and installation process.  Finally, using Chelsio’s high-speed Unified Wire concept, Purdue is able to collect unused cycles across the system to deliver a true supercomputing resource for other scientific work – achieving extremely high levels of system utilization and cluster efficiency.

Chelsio’s offload technology was further able to deliver power and compute cycle savings.  At Purdue’s target of 95% utilization, this amounts to fewer network fabrics within the cluster, a significant reduction in TCO, and an increase in the effective available compute cycles for the same investment dollars.

“We had a stringent requirement list, led by the need for RDMA Clustering, and only Ethernet and Chelsio’s proven iWARP implementation were able to fulfill this need.  This cluster is expected to last five years and fulfill disparate requirements of various departments in the University,” said John P. Campbell, associate VP of the Rosen Center for Advanced Computing, Purdue University.   “The Coates Cluster is expected to place well in the Top-500 list later this year.”

“With the multi-vendor IETF standard iWARP protocol, Ethernet RDMA clustering is especially suited for the Coates Cluster.  Purdue University recognizes the many benefits of converging its LAN, storage and HPC traffic over offloaded 10Gb Ethernet, which results in a scalable ultra-high performance network at a much lower TCO,” said Kianoosh Naghshineh, president and CEO of Chelsio.  “We are very pleased to be able to be an active participant in the world’s largest 10GbE cluster which concurrently is the largest iWARP installation as well.”