Chapter 6. Multi-rack Considerations
If you are building a cluster consisting of multiple racks (common for large clusters), consider utilizing as much network bandwidth between switches in a "fat tree" design for optimal performance. A typical 10Gbps Ethernet switch has 48 10Gbps ports and four 40Gbps ports. If you only use one one 40Gbps port for connectivity, you can only connect 4 servers at full speed (i.e., 10gbps x 4). Use your 40Gbps ports for maximum throughput. If you have unused 10G ports, you can aggregate them (with QSFP+ to 4x SFP+ cables) into more 40G ports to connect to other racks and to spine routers.
For network optimization, we recommend a jumbo frame for a better CPU/bandwidth ratio. We also recommend a non-blocking network switch back-plane.