Beta

Hypervisor Scale

VMs that outperform bare metal.

Hypervisor Scale

The Hypervisor Scale is the foundation of Tophan’s compute layer. It manages virtual machines with a focus on performance, density, and operational simplicity.

Why It’s Different

Most virtualisation platforms treat performance as an afterthought — they add layers of abstraction and hope the overhead stays acceptable. Tophan’s Hypervisor Scale is built from the ground up around KVM, with memory deduplication and intelligent scheduling that regularly lets VMs outperform equivalent bare-metal configurations through better resource utilisation.

Features

FeatureDescriptionStatus
KVM IntegrationDirect kernel-level virtualisation. No middleware, no translation layers.Stable
KSM Memory DedupIdentical memory pages across VMs are stored once. 100 identical VMs use barely more RAM than 10.Stable
GPU PassthroughFull PCI passthrough for GPUs. Native driver performance inside VMs for AI/ML, rendering, and desktop workloads.Beta
QMP ProtocolMachine-level VM control via QEMU Machine Protocol. Scriptable, automatable, no UI required.Stable
Live MigrationMove running VMs between nodes with zero downtime. Millisecond switchover.Beta
VM TemplatesGolden images for rapid provisioning. One template, unlimited deployments, all deduplicated.Beta
Snapshot ManagementPoint-in-time snapshots with instant rollback. Layered on the immutable storage model.Beta
Resource LimitsCPU pinning, NUMA-aware placement, memory ballooning, I/O throttling.Stable
SPICE ConsoleFull graphical console access through Dragon’s Eye. No separate client needed.Beta
Cloud-InitAutomatic VM configuration on first boot. Networking, users, packages, scripts.Stable

Architecture

┌──────────────────────────┐
│      Dragon's Eye        │  Management
├──────────────────────────┤
│       tophand            │  Orchestration
├──────────────────────────┤
│   Hypervisor Scale API   │  VM lifecycle
├──────────────────────────┤
│     QMP / libvirt        │  Machine control
├──────────────────────────┤
│        KVM / QEMU        │  Virtualisation
├──────────────────────────┤
│     Linux Kernel         │  Hardware access
└──────────────────────────┘

The Hypervisor Scale exposes a clean REST API consumed by tophand. All VM operations — create, start, stop, migrate, snapshot — flow through this API. Dragon’s Eye provides the human interface, but every operation is equally accessible via CLI or automation.

Memory Deduplication in Practice

KSM (Kernel Same-page Merging) is always active. When you run 50 VMs from the same template, the hypervisor detects identical memory pages and stores them once. As VMs diverge, only the differences consume additional RAM.

This means you can dramatically overprovision memory — safely — because real consumption tracks actual divergence, not nominal allocation. A host with 128GB of RAM can comfortably run workloads with 300GB+ of nominal allocation if they share a common base.

GPU Passthrough

Full IOMMU-based PCI passthrough gives VMs direct hardware access to GPUs. The VM runs the manufacturer’s native drivers with no performance penalty. This is essential for:

  • AI/ML training and inference workloads
  • GPU-accelerated desktop VMs (via the Desktop Scale)
  • Rendering and video encoding pipelines