4.4. Multi-Rail Topologies
A multi-rail configuration provides load balancing and failover capabilities, adding a higher degree of fabric redundancy. If one SuperNIC or an entire subnet fails, traffic can be moved to the remaining switches.
The multi-rail feature can be applied to a single subnet or multiple subnets. By enabling multi-rail, a process can use multiple network interface cards (SuperNICs) to transfer messages.
Three basic scenarios include:
Single-rail in a single subnet: This scenario consists of one SuperNIC in a server connected to one subnet. This is the default configuration during installation. This configuration provides the performance required by most applications in use today.
Dual-rail in a single subnet: This scenario consists of two SuperNICs in the same server connected to the same subnet. This configuration also provides improved MPI message rate, latency, and bandwidth to the node.
Dual-rail in dual subnets: This scenario consists of two SuperNICs in the same server connected to separate subnets. Depending on the platform, this configuration may provide improved MPI message rate, latency, and bandwidth to the node. This configuration is also commonly referred to as a ‘dual-plane’ configuration, as the two subnets exists as individual topological planes.
Note
Other multi-rail scenarios can be configured. A single Host Fabric Manager (FM) server can manage multiple subnets up to the supported number of 0-7 Fabric Manager instances.