More Information...

White Paper: FocalPoint in Large Scale Clos Switches (PDF)

Using FocalPoint in High-Performance Computing

The extremely low latency of the FocalPoint FM4000 family make it ideal for computer clustering applications such as High Performance Computing (HPC). FocalPoint switches also provide scalability with up to 24 ports of 10 Gigabit Ethernet connectivity. If higher bandwidth connections are needed, Ethernet’s native Link Aggregation Protocol can be used to aggregate multiple 10 Gigabit ports into a single, higher bandwidth pipe of up to 160Gb/s. With FocalPoint, link aggregation can be done even across multiple switches in a multi-stage topology such as a fat tree or stack while maintaining non-blocking, full bandwidth operation.

In dense installations, compute blade shelves are used which typically contain slots for Ethernet switch modules. By using FM4224 FocalPoint switches in these modules, 12 10G ports can be used to locally interconnect the compute blades while the other 12 ports can be used as as external uplinks to Top of Rack (ToR) switches. By using two switch modules per shelf, up to 24 compute blades can be implemented on a single shelf.

In the example below, the switch module uplinks are connected to twelve 24-port ToR switches. In this way, up to 12 compute blade shelves can be interconnected through these switches. FocalPoint provides features to automatically load balance traffic across the ToR switches creating and effective 120G data pipe. If a ToR switch fails, data can be automatically load balanced around the failed device providing N+1 redundancy.

With a single 10G Ethernet connection per blade, up to 288 compute blades can be interconnected in this fashion. With redundant 10G Ethernet connections to each blade, up to 144 compute blades can be supported. Finally the switches exhibit an extremely low latency of only 200 nanoseconds for layer 2 switching and 300 nanoseconds for layer 3 routing. Using FocalPoint switches, scalable high-performance computing can be achieved with less than 600nS blade-to-blade latency.