Experiment with Cloud Hypervisor differential snapshots

[sjmiller609]

summary

• add a feature-gated Cloud Hypervisor differential snapshot path
• retain an immutable memory baseline and merge sparse deltas with range reflinks
• detach retained memory before merging so fanout siblings and the running VM keep the prior view
• expose the snapshot mode and restore-time dirty-tracking fields used by the corresponding Cloud Hypervisor v51.1 changes

validation

• go test ./lib/hypervisor/cloudhypervisor
• targeted snapshot ownership tests in ./lib/instances
• go test ./lib/vmm on KVM-enabled Linux
• Cloud Hypervisor: 56 VMM unit tests and cargo check -p cloud-hypervisor
• five consecutive Chromium standby/restore cycles succeeded
• 25-way Chromium fanout, concurrent differential standby, and 25/25 restore validation succeeded

benchmark

For 25 simultaneous 8 GiB Chromium standbys, the v51.1 fork with 64 KiB dirty-range coalescing:

• completed 25/25 in 43.1s (42.4s p95)
• wrote 7.34 GB of snapshot delta data; device-level writes were 8.39 GB
• produced 34,419 extents total

A full snapshot writes 8.60 GB per VM, or about 215 GB for 25 VMs. The differential run reduced device-level writes by about 96%. Coalescing reduced the extent count by 85% and wall time by 41% versus the uncoalesced differential run, though the two fresh runs had somewhat different dirty-page totals.

status

This remains behind HYPEMAN_EXPERIMENTAL_CH_DIFF_SNAPSHOTS and requires the corresponding Cloud Hypervisor v51.1 fork. It is intended for benchmarking and design validation, not rollout. Production hardening still needs per-instance eligibility enforcement, crash-atomic baseline publication, a controlled dirty-logging overhead benchmark, and an owned build/release path for the fork.