Network Architectures Compared: Centralized, Cluster, Cloud and Virtualized Networks

Students often meet the terms centralized, cluster, cloud and virtualized network in one exam question and assume they are four sizes of net...

Students often meet the terms centralized, cluster, cloud and virtualized network in one exam question and assume they are four sizes of network, like LAN and WAN. They are not. They describe how services and network functions are arranged. In this post I put all four side by side, separate cloud networking from cloud computing (the most common mix-up), and show VLAN, VXLAN and SDN as concrete examples of virtualization you will configure in your career.

Centralized network

In a centralized architecture, one central system (a server, a server room or a single data center) does the important processing and stores the data, and every user connects to it. Clients are relatively simple; they send requests and display results.

Classic examples are a mainframe with terminals, a bank's core banking system that every branch connects to, and many messaging services where all messages pass through the provider's servers. In Nepal, a typical setup is a company head office in Kathmandu hosting the ERP and file servers, with branches in Pokhara and Butwal reaching it over a VPN or MPLS link.

Strengths

• Simple control. Users, permissions, patches and logs are managed in one place.
• Consistent data. There is one copy of the truth, so no syncing between sites.
• Easy monitoring. Because traffic flows to one point, it is easy to log and inspect.
• Users can move. If one PC fails, the user logs in from another and their data is still on the server.

Weaknesses

• Single point of failure. If the central server or the link to it goes down, every user is affected unless you add redundancy.
• Scaling up is limited. Adding CPU, RAM, disk or bandwidth to one box (vertical scaling) eventually hits a ceiling. Beyond that you must add more servers and load balancing, which is no longer purely centralized.
• WAN dependence. Remote branches are only as fast as their link to the center.
• Security is concentrated. One place to protect is easier to manage, but it is also one attractive target. Centralized is not automatically "more secure"; it depends on how well that center is defended.

Cluster network

A cluster is a group of servers (nodes) that work together so clients see them as one system. Strictly speaking a cluster is a computing architecture; the "cluster network" is the network that ties the nodes together and to users. The nodes are usually in the same data center, connected by fast LAN links, and often have a separate private link for heartbeat and replication traffic.

Three kinds of cluster

• High-availability (failover) cluster. One node is active, another stands by. The nodes exchange heartbeat messages; if the active node stops responding, the standby takes over its IP address and workload. Windows Server Failover Clustering and database clusters work this way. Network engineers see the same idea in firewall HA pairs and in HSRP/VRRP gateway redundancy.
• Load-balancing cluster. Several nodes serve requests at the same time, with a load balancer spreading traffic between them. Web server farms are the usual example.
• High-performance computing (HPC) cluster. Many nodes split one large calculation, such as weather modeling or scientific research, and need very low-latency links between them.

A cluster differs from grid computing mainly in tightness: cluster nodes are usually similar machines in one location under one administrator, while a grid joins loosely connected, often different machines across sites.

Strengths and weaknesses

Clusters improve availability (one node can fail without an outage) and, for load-balancing and HPC clusters, capacity. They do not automatically make a single task faster; that depends on whether the application can be spread across nodes. The costs are real: more hardware, shared storage or replication, cluster licensing, and more complex monitoring. A badly designed cluster can even fail completely if nodes lose contact and both try to become active (a "split-brain"), which is why clusters use a quorum or witness to decide who wins.

Cloud network vs cloud computing

This is where most notes go wrong, so let's separate the two ideas clearly.

Cloud computing: what you rent

Cloud computing is renting computing resources from a provider over a network, on demand, and paying for what you use. NIST SP 800-145 defines the widely used terms:

• Service models (how much the provider manages): IaaS (you get virtual machines, storage and networks, e.g. Amazon EC2), PaaS (you deploy code to a managed platform), SaaS (you just use the finished application, e.g. Gmail or Microsoft 365).
• Deployment models (who the cloud is for): public, private, hybrid and community.

Public, private and hybrid are therefore types of cloud deployment, not types of network.

Cloud networking: how it is connected

Cloud networking is about the network itself, and it covers two different situations:

1. Networks built inside a cloud provider. In AWS, Azure or Google Cloud you create a virtual private network (called a VPC or VNet), divide it into subnets, write route tables, attach security groups, and add NAT gateways, load balancers and VPN or dedicated links back to your office. None of this is physical to you; it is software running on the provider's infrastructure.
2. Cloud-managed networks. Your switches, access points and firewalls are physically in your office, but they are configured and monitored from a vendor's cloud dashboard (Cisco Meraki is a well-known example). The data stays on your LAN; only management goes to the cloud.

A Nepali software company hosting its app on AWS in the Mumbai region, with an IPsec VPN from its Kathmandu office to its VPC, is using both: cloud computing (IaaS) and cloud networking (VPC, subnets, VPN).

Trade-offs

Benefits: no large upfront hardware purchase (you move spending from capital to operating costs), fast scaling, global regions and built-in redundancy options. Costs: everything depends on your internet link, monthly bills can grow quickly if not watched, data leaving the cloud (egress) is often charged, and security becomes a shared responsibility: the provider secures the platform, you secure your configuration. Misconfigured security groups and public storage buckets are a frequent cause of real-world breaches.

Virtualized network

Network virtualization means creating logical networks in software on top of shared physical hardware, so several isolated networks can run on the same switches and cables. Do not confuse it with server virtualization (many virtual machines on one physical host). They often appear together, but they solve different problems.

VLAN: virtualization you already know

A VLAN (IEEE 802.1Q) splits one physical switch into several separate Layer 2 broadcast domains. Trunk links carry many VLANs by inserting a 4-byte tag with a 12-bit VLAN ID into each frame, giving IDs 1 to 4094. A small office might put staff on VLAN 10, guests on VLAN 20 and cameras on VLAN 30, all on the same switches.

Switch(config)# vlan 20
Switch(config-vlan)# name GUEST
Switch(config-vlan)# exit
Switch(config)# interface fastEthernet 0/5
Switch(config-if)# switchport mode access
Switch(config-if)# switchport access vlan 20

switchport mode access makes the port carry one untagged VLAN, and switchport access vlan 20 places the device on that port in the guest network.

VXLAN: VLANs for data centers

4094 VLANs is not enough for large data centers and cloud providers with thousands of tenants, and plain VLANs cannot stretch across a routed network. VXLAN (RFC 7348) solves both. It wraps the original Ethernet frame inside a UDP packet (destination port 4789) and uses a 24-bit VXLAN Network Identifier, allowing about 16 million segments. Devices called VTEPs do the encapsulation, and the frames travel across a normal routed IP "underlay". Cloud VPCs are built on overlay techniques of this kind.

Figure: How a VTEP wraps an Ethernet frame for transport across the IP underlay

SDN: separating the brain from the muscle

In a traditional network each router or switch has its own control plane (deciding where traffic goes) and data plane (actually forwarding it). Software-Defined Networking moves the control decisions to a central controller that programs many devices through APIs, using protocols such as OpenFlow or vendor APIs. Cisco Catalyst Center (formerly DNA Center), Cisco ACI and SD-WAN controllers are examples. Related is NFV (Network Functions Virtualization): running firewalls, routers and load balancers as virtual machines instead of dedicated boxes.

Strengths and weaknesses

Virtualization lets you create, change and isolate networks in minutes, use hardware more efficiently and automate repetitive work. In return it needs engineers who understand both networking and the virtualization layer, adds layers that make troubleshooting harder, and can be costly to license and design properly at the start. Most organizations accept these costs because the long-term savings in hardware and time are significant.

Comparison table

Architecture	Core idea	Typical example	Main strength	Main weakness
Centralized	One central system serves all users	Head office ERP accessed by all branches	Simple control and consistent data	Single point of failure; vertical scaling limits
Cluster	Several servers act as one system	Failover database pair; web server farm	High availability and added capacity	Cost and complexity; split-brain risk
Cloud	Rent compute and networks from a provider; build virtual networks there	App on AWS with a VPC and VPN to the office	Fast scaling, no upfront hardware	Internet dependence, ongoing cost, misconfiguration risk
Virtualized	Logical networks in software on shared hardware	VLANs, VXLAN overlays, SDN controllers	Flexibility, isolation, automation	Skills needed; harder troubleshooting

These architectures combine freely. A cloud provider runs clusters of servers, connected by virtualized (VXLAN and SDN) networks, while your company may still use it in a centralized way.

Prefer video? Watch more hands-on labs on the Technical Hakim YouTube channel.

Summary

• Centralized: one center does the work. Easy to manage, but plan for its failure.
• Cluster: many servers acting as one, for availability or capacity.
• Cloud computing is what you rent (IaaS, PaaS, SaaS; public, private, hybrid, community). Cloud networking is how those resources are connected and managed.
• Network virtualization: VLAN at the small scale, VXLAN for data centers, SDN for central control.

Related tutorials

• Configuring Static VLANs on a Cisco Switch (Packet Tracer Lab)
• Trunking and 802.1Q on Cisco Switches: Configs and Common Errors
• HSRP Configuration for Gateway Redundancy (Packet Tracer Lab)
• Site-to-Site IPsec VPN on Cisco Routers: Packet Tracer Lab
• Types of Computer Networks: LAN, MAN, WAN, PAN (and more) Explained

Sources

• NIST SP 800-145, The NIST Definition of Cloud Computing
• IEEE 802.1Q, Bridges and Bridged Networks (VLAN tagging)
• RFC 7348, Virtual eXtensible Local Area Network (VXLAN): https://www.rfc-editor.org/rfc/rfc7348
• RFC 7426, Software-Defined Networking (SDN): Layers and Architecture Terminology: https://www.rfc-editor.org/rfc/rfc7426
• Cisco IOS Configuration Guide – VLANs