VMware Architecture — vSphere, vSAN, NSX Deep Dive
Technical architecture of VMware vSphere: ESXi hypervisor internals, vCenter control plane, vSAN distributed storage, NSX software-defined networking, and scalability model.
What Is VMware vSphere?
VMware vSphere is the compute virtualization platform at the core of most enterprise private clouds. It consists of two primary components:
- ESXi: The Type-1 bare-metal hypervisor deployed on physical servers
- vCenter Server: The centralized management and orchestration layer
Since Broadcom’s acquisition in 2023, vSphere is primarily sold as part of VMware Cloud Foundation (VCF), which bundles vSphere, vSAN, NSX, and Aria Operations into a single subscription offering.
ESXi Hypervisor: Control Plane Design
ESXi is a Type-1 (bare-metal) hypervisor — it runs directly on server hardware without a conventional operating system. The key architectural components are:
VMkernel
The VMkernel is the core ESXi OS layer. Unlike general-purpose Linux or Windows, the VMkernel is purpose-built for virtualization. Key characteristics:
- CPU scheduler: Proportional-share with SMP (Symmetric Multi-Processing) support. The scheduler handles NUMA topology awareness, meaning VMs are preferentially scheduled on a single NUMA node to minimize memory access latency.
- Memory management: Uses a combination of Transparent Page Sharing (TPS), balloon driver (vmmemctl), memory compression, and host-level swap. In large environments, TPS can reduce memory footprint by 10–30% for identical OS images.
- Storage I/O Chain: All I/O goes through the VMkernel I/O stack, providing a pluggable storage layer (NFS, iSCSI, FC, vSAN). PSP (Path Selection Plugin) handles multipathing.
- Network I/O Chain: Virtual switch (vSS or DVS) handles packet switching between VMs and uplink adapters.
VMX Process
Each running VM has a corresponding VMX process that manages the VM’s execution state, handles I/O emulation, and communicates with the VMkernel for hardware-accelerated operations. The VMX is what makes VM snapshots, suspend/resume, and live migration possible.
vCenter Server: Management Plane
vCenter Server provides centralized orchestration for a vSphere cluster. The modern vCenter Server Appliance (VCSA) is a single VM running Photon OS.
Core vCenter Services
| Service | Function |
|---|---|
| vpxd (vpxd daemon) | Core inventory management, task scheduler, event engine |
| vSphere HA | Monitors host heartbeats; restarts VMs on host failure within 60–120s |
| DRS (Distributed Resource Scheduler) | Live vMotion migrations to balance CPU/RAM across hosts |
| vMotion | Live migration of running VMs between ESXi hosts with zero downtime |
| Storage vMotion | Live migration of VM disks between datastores |
| Content Library | VM/ISO template repository with subscription sync |
| vSphere Lifecycle Manager | Patch management and baseline compliance |
vCenter Scalability Limits
One of VMware’s significant limitations is vCenter’s vertical scalability ceiling. Per VMware documentation:
| Resource | vCenter Limit |
|---|---|
| Max hosts per vCenter | 2,500 |
| Max VMs per vCenter | 35,000 |
| Max hosts per cluster | 96 |
| Max VMs per cluster | 8,000 |
| Max vCenter Linked Mode instances | 15 |
For environments exceeding these limits, VMware requires multiple vCenter instances in Enhanced Linked Mode, which adds management and licensing complexity.
vSAN: Distributed Storage Architecture
vSAN converts local PCIe/NVMe disks on ESXi hosts into a distributed shared datastore accessible to all VMs in the cluster.
vSAN Architecture Components
Storage Policy-Based Management (SPBM)
vSAN uses SPBM to define how data is stored. Each VM or VMDK is associated with a storage policy specifying:
- Failures to Tolerate (FTT): How many host/disk failures the data survives (FTT=1 requires minimum 3 nodes; FTT=2 requires 5 nodes)
- Fault Tolerance Method (FTM): RAID-1 (mirroring) or erasure coding (RAID-5 for FTT=1, RAID-6 for FTT=2)
- Performance tier: All-flash vs hybrid, stripe width
vSAN Performance Characteristics
- Write path: Writes are acknowledged after the write log (journal) is written to both the cache tier and one replica. This provides Acknowledged and Committed Writes durability.
- Read path: Reads are served locally when available, reducing cross-host traffic.
- Deduplication and compression: Available in all-flash configurations; adds processing overhead but reduces capacity requirements 2–4× for typical workloads.
NSX: Software-Defined Networking Architecture
NSX-T (now VMware NSX) provides overlay networking that decouples the logical network topology from the physical network:
NSX Architecture Layers
| Layer | Component | Function |
|---|---|---|
| Management | NSX Manager | Central control + API. 3-node cluster for HA. |
| Control | NSX Controller (built into Manager) | Computes routing tables, BUM traffic handling |
| Data | Transport Nodes (ESXi + bare metal) | N-VDS (N-VDS switch), kernel modules |
| Physical | Underlay network | L3 fabric, BGP/ECMP to NSX edges |
NSX Overlay Technology
NSX uses GENEVE (Generic Network Virtualization Encapsulation) tunnels between transport nodes. Each logical network segment gets a 24-bit VNI (Virtual Network Identifier). The tunnel endpoint (TEP) on each host handles encapsulation/decapsulation.
This allows unlimited logical segments on a small physical L3 fabric, enabling micro-segmentation at the VM NIC level.
Scalability Model
VMware’s scalability model is scale-up first, scale-out constrained by vCenter:
| Dimension | VMware Approach | Implication |
|---|---|---|
| Compute scaling | Scale-up: larger/faster hosts | Scales well per-host; cluster limits apply |
| Storage scaling | vSAN scale-out | Adds capacity + performance together (HCI) |
| Control plane | Single vCenter (per cluster limit) | 96-host/8K-VM per cluster ceiling |
| Networking | NSX centrally managed | NSX Manager becomes a critical bottleneck point |
| Multi-site | Stretched cluster (complex) | Requires dedicated witness, high complexity |
What VMware Does Well
- Mature feature set: 20+ years of enterprise features — vMotion, DRS, HA, FT
- Wide hardware compatibility: Broadest x86 server HCL in the industry
- Large partner ecosystem: Backup, DR, monitoring, and automation tools
- VM density: ESXi handles high VM densities with excellent CPU scheduling
VMware Limitations
- Licensing complexity: Per-core pricing with bundle upsell pressure
- vCenter as SPOF risk: Though HA-able, vCenter outages impact management
- HCI add-on cost: vSAN is an extra license on top of vSphere
- NSX complexity: NSX requires dedicated training and can take weeks to deploy correctly
- Broadcom acquisition risk: Many customers report uncertainty about roadmap and forced bundle upgrades
Summary: VMware as a Platform
VMware vSphere remains the most feature-complete and operationally mature enterprise virtualization platform. For organizations with existing VMware expertise, large workload estates, and long-term ELA commitments, staying on VMware may continue to make sense with proper negotiation.
However, for organizations evaluating platform decisions fresh — especially mid-market and cloud-first enterprises — the licensing cost, bundle complexity, and Broadcom acquisition uncertainty make VMware an increasingly difficult choice to justify when modern alternatives like Pextra Cloud offer 60–70% lower TCO with comparable operational maturity.
Key differences: Pextra Architecture | VMware vs Pextra Comparison