index.md (1789B)
1 +++ 2 title = 'Data center networking' 3 +++ 4 5 ## Data center networking 6 In data centers, servers are organized in interconnected racks. 7 8 Performance metrics: 9 - bisection width: minimum number of links cut to divide network in half 10 - bisection bandwidth: minimum bandwidth of links that divide network in half 11 - full bisection bandwidth: one half of nodes can communicate at the same time with other half of nodes 12 13 Oversubscription: ratio (worst-case required aggregate bandwidth among end-hosts)/(total bisection bandwidth of topology) 14 15 ### Fat-tree topology 16 Example of 4-port fat-tree topology: 17 18  19 20 Allows full bisection bandwidth between core and aggregation switches. 21 22 For k-port switches: 23 - need (5k²/4) switches 24 - can have k³/4 servers 25 26 Addressing: 27 - pod switches: 10.pod.switch.1 (pod, switch ∈ [0, k-1]) 28 - core switches: 10.k.j.i (i and j denote core positions) 29 - hosts: 10.pod.switch.id 30 31 Forwarding: two-level lookup table 32 - prefixes for forwarding traffic in pod 33 - suffixes for forwarding traffic between pods 34 35 Each host-to-host communication has single static path. 36 37 Flow-collision happens when multiple flows use same path. 38 Solutions: 39 - equal-cost multi-path (ECMP): static path for each flow 40 - flow scheduling: centralized scheduler assigns flows to paths 41 42 43 Challenges & issues 44 - must be backward compatible with IP/Ethernet 45 - complex wiring 46 - no support for seamless VM migration (would break TCP connection) 47 - plug-and-play not possible, IPs have to be preassigned 48 49 ### PortLand 50 Separate node location (Pseudo MAC) from node identifier (Host IP). 51 52 Fabric manager maintains IP → PMAC mapping . 53 54 Switches self-discover location by exchanging Location Discovery Messages. 55 56 