Applications

• Geological analysis for oil and gas exploration
• Bioscience and genome mapping
• Nanotechnology research and development
• Financial analysis
• Fluid dynamics
• DoD testing and simulation
• Weather forecasting
• Data mining / predictive optimization

Computer clusters have provided cost-effective alternatives to traditional supercomputers for thousands of universities, governments, research institutions and companies. Many scientific and other numerical-computation problems can be decomposed to run concurrently over low-cost computers.

Today, each node in computer clusters commonly relies on multiple networks that serve different purposes.

The LAN lets the users operate and control the cluster, and reaches the NAS shared file-storage system. Since the LAN links the clusterto the outside world, it needs to be a standard and widely-deployed technology. Today, Gigabit Ethernet is common.

The cluster fabric interconnects the cluster nodes, conveying the inter-processor communications (IPC) messages that are key to dividing applications’ algorithms among multiple nodes. Applications use message-passing middleware such as MPIto provide a programming interface for cluster applications. Good performance requires cluster fabrics to minimize message-passing latency. Today, Gigabit Ethernet is a commonly-used cluster fabric, and Myrinet and InfiniBand are used when latency requirements are more stringent.

Many clusters use a storage area network (SAN) to connect the cluster nodes to a shared block-level storage system. These networks require high performance with low protocol-processing overhead on the processors in the servers and storage equipment. Today, the most common SAN fabric is Fibre Channel, available only in 1, 2, and 4Gbps versions.

Thus, each server in the cluster participates in as many as three different networks – the LAN, clusterfabric, and SAN . Each network requires one or more switches, its own management tools and network adapters for each server. In addition to the copper cabling for Gigabit Ethernet, the overlay networks for the clusterfabric and LAN each requires expensive fiber cable connections for each server. Each technology requires its own technical skills and inventory of spare components to deploy and maintain.

Enter10 Gigabit Ethernet
10 Gigabit Ethernet is the driver forHPCC toprovidedramatically m ore at less cost.

In the next generation cluster, a single 10 Gigabit Ethernet network can carry all the LAN, SAN, and cluster traffic. 10GbE increases bandwidth and shrinks latencies. With the acceleration provided by Chelsio’s protocol-offload adapters, 10GbE offers sub-10 microsecond message latency. Further, in servers and storage systems where protocol processing is offloaded to these Chelsio protocol engines, CPU overhead is very low.

Ethernet Cost Profile -Ethernet is the most widely deployed networking technology. Each Ethernet generation’s scale economies have driven down prices farther and faster than any other networking technology. Today, 10GbE largely uses fiber media. Chelsio also offers the 10GbE network interface for the CX4 copper cabling standard along with the twisted copper pair wiring for 10GBase-T interfaces at a low cost.

Benefits Summary
10 Gigabit Ethernet with Chelsio’s protocol offload technology provides dramatic advantages for HPCC.

• High bandwidth
• Low latency
• Low communications overhead against application performance
• Facilitates converging to single network technology
• Leverages Ethernet’s tremendous scale economies to drive down prices