architecture.md 3.3 KB
title: Network Architecture breadcrumbs:
- title: Networking --- {% include header.md %}
Topologies
Flat
- A single L2 domain/layer.
- Switching, routing and firewalling is all done by the same device or device stack, with clients directly connected.
Three-layer Hierarchical Model
- Appropriate for large networks spanning multiple regions (e.g. multiple buildings).
- Scales well.
- Focuses on north-south traffic.
- Consists of three layers.
- Access layer:
- L2 switches.
- Connected to clients.
- Typically one access-layer VLAN spans one or a few access switches.
- Should implement first-hop security (MAC, IPv4, IPv6).
- Connected upstream to distribution switches.
- Distribution layer:
- Aka "distro" layer.
- L3 switches.
- Terminates access-layer VLANs.
- Implements features like filtering and QoS.
- May manage individual WAN connections.
- Connected upstream to core routers and ptionally interconnected with other distribution switches.
- Core layer:
- Routers.
- Provides a backbone between distribution regions and toward external networks.
- Focuses entirely on high bandwidth, low latency, high reliability and high resilience.
- Avoids anything that may slow down traffic, like access lists, policy enforcement, etc.
- It's possible to connect multiple core routers and distribution switches by using a switch.
Collapsed Core
- Similar to the three-layer hierarchical model, but with the core and distribution layers collapsed into the same devices.
- This generally means that there is only one routed layer.
- Distro layer devices may be interconnected directly or through one or more core switches (not routers) which are not themselves interconnected.
- Appropriate for medium/small sites without multiples regions, where a separate core network is not needed.
Collapsed Distribution
- Similar to the three-layer hierarchical model, but with the distribution and access layers collapsed.
- Generally not very useful.
Spine-Leaf
- A type of Clos network (non-hierarchical).
- Two or three layers: Leaf layer, spine layer and an optional super-spine layer (for larger networks).
- Leaf switches connect to every spine switch and not to any other leaf switches.
- Spine routers (or switches) are not connected to any other spine routers.
- Hosts connect only to leaf switches.
- All spine-leaf links are L3 (routed).
- Every pair of leaf switches are always two hops away from each other.
- Routers to external areas, firewalls and load balancers are added connected leaf switches called border leaves.
- Large spine-leaf networks may be broken into multiple networks where the spine rouers are connected to routers in the super-spine layer.
- Focuses on east-west traffic.
- Requires ECMP for optimal utilization.
- Well suited for data centers.
- Well suited for VXLAN for allowing hosts to move easily between leaf switches.
Terms
- Equal-cost multi-path routing (ECMP): Routing strategy for forwarding over multiple best paths to the same destination.
- Oversubscription: Less uplink capacity than downlink capacity. Appropriate when downstream devices rarely use the uplink simultaneously, allowing them to share the uplink capacity without sacrifice.
Miscellanea
- Q-in-Q (simple) or VXLAN (sophisticated) may be used to span VLANs over different areas.
{% include footer.md %}