QEMU Snaphosts Are So Slow. Really?
summary
Implementation adds a new snapshot data path with a rewritten cluster-aligned IO engine, delegated read worker threads, and a new on-disk snapshot format; result is raw reads climb from near-zero to about 150 MB/s observed on fast disks while rotational disks still show very poor results. Snapshot creation uses background indexing and a notification engine so active guests pause briefly — measured at ~75 ms downtime — and most patches are functional and ready to merge after corner-case fixes for memory-indexing and merge-on-write.
AI-written summary · about
chapters (11)
- 0:00QEMU snapshots performance
- 1:10Motivation for faster snapshots
- 3:00Current QEMU snapshot implementation
- 6:24Problems with standard formats
- 8:52Block driver challenges
- 11:04Performance with rotational disks
- 13:20Cluster alignment for performance
- 15:09Memory protection during snapshots
- 17:02Active snapshot limitations
- 19:56New snapshot format challenges
- 21:27Conclusion and future
related talks
20mBackground Snapshots in QEMU: Towards Asynchronous Revert
21mNVMe Emulation Performance Optimization
25mXen and XenServer Storage Performance
25mKVM Dirty Page Tracking
11mNew Qemu Backup Architecture and API
40mAchieving the Ultimate Performance with KVM
35mIntroduce an SPDK vhost FUSE Target to Accelerate File Access in...
26mAnalysis of AMD HW-assisted vIOMMU Implementation and Performance