How does live migration work at hardware layer?
In a virtualized world, Live Migration (moving a running VM from one physical server to another with zero downtime) feels like magic. To the end-user, it’s a seamless transition. But at the hardware layer, it’s a high-stakes, high-speed "relay race" between CPUs, Memory controllers, and Network Interface Cards.
Here is how the physical hardware pulls off the feat of moving a "living" workload.
The biggest challenge of live migration is the RAM. A running database or application has gigabytes of data stored in the physical memory (DRAM) of Source Server A.
The Hardware Action: The Hypervisor uses the MMU (Memory Management Unit) to mark all memory pages of the VM as "Read-Only."
The Transfer: While the VM is still running on Server A, the hardware begins copying all memory pages to the DRAM of Server B via the high-speed interconnect (RoCE or InfiniBand).
While the memory is being copied, the VM is still "alive" and writing new data. These changed areas are called Dirty Pages.
The Hardware Action: Modern CPUs use a feature called PML (Page Modification Logging). Instead of the software hypervisor having to scan all memory to find changes, the CPU hardware maintains a log of every memory address that was modified.
The Iteration: The system sends these "dirty" pages to the target server. This happens in multiple rounds, each round getting faster as the volume of changed data shrinks.
Once the remaining "dirty" pages are small enough to be sent in a few milliseconds, the hypervisor triggers the Suspend phase.
The Hardware Action: The CPU registers, instruction pointers, and interrupt states are "snapshotted."
Context Transfer: This tiny packet of data (the CPU's current thought process) is sent to Server B.
Hardware Compatibility: For this to work, the CPUs must be "Instruction Set Compatible." You generally cannot live-migrate a VM from an Intel Sapphire Rapids CPU to an older Broadwell chip because the hardware "vocabulary" (instructions) is different.
At this exact moment, the VM is physically located on Server B, but the rest of the network still thinks it's on Server A.
The Hardware Action: The Network Interface Card (NIC) on the new server immediately broadcasts a Gratuitous ARP (Address Resolution Protocol) packet.
The Switch Update: This hardware-level broadcast tells the physical network switches: "IP 10.0.0.5 is no longer at Port 1; it is now at Port 24!"
The Result: Network traffic is rerouted in microseconds, often before a single TCP packet times out.
For a "Live" migration to work at the hardware layer, the actual data (the Virtual Disk) does not move. Both physical servers must be connected to a Shared Storage Fabric (like a SAN or Exadata Storage Grid). The hardware simply passes the "ownership" of the disk locks from the HBA (Host Bus Adapter) of Server A to the HBA of Server B.
Modern servers use specialized hardware to make this faster:
| Technology | Hardware Role |
| RDMA (RoCE) | Moves memory from Server A to Server B without involving the CPU, reducing migration time by up to 80%. |
| AES-NI | If the migration stream is encrypted, the CPU uses hardware-accelerated instructions to encrypt the RAM data at line speed. |
| IOMMU | Manages the "re-mapping" of virtual I/O devices so the VM can find its virtual disk on the new physical bus. |
Live migration is a coordinated dance of hardware. The RAM is mirrored, the CPU state is teleported, and the Network Switch is updated. By the time the guest OS realizes something has happened, it is already executing its next instruction on a completely different piece of silicon.