Performance: Under the Hood

Companion to Performance.

Boot time: machine type

The machine type affects boot time:

Machine	Boot time	Why
q35 / virt	~2-3 seconds	Full chipset emulation, firmware device scan

Direct kernel boot (see below) is the primary boot-time optimization, skipping BIOS/UEFI entirely.

Boot time: direct kernel boot

-kernel /path/to/vmlinuz \
-initrd /path/to/initrd.img \
-append "root=/dev/vda rw rootfstype=ext4 quiet loglevel=0 ..."

Passing the kernel directly to QEMU skips BIOS/UEFI firmware entirely. Combined with quiet loglevel=0, the kernel suppresses most boot messages, and raid=noautodetect skips MD RAID scanning.

CPU acceleration

# Hardware acceleration (near-native):
-accel kvm -cpu host    # Linux
-accel hvf -cpu host    # macOS
-accel whpx -cpu host   # Windows

# Software emulation (10-50x slower):
-accel tcg -cpu max

-cpu host exposes the real CPU features to the guest. The hypervisor runs guest code directly on the hardware. -cpu max (TCG) emulates the most capable CPU QEMU supports, but every instruction is translated in software.

Disk I/O

-object iothread,id=iothread0 \
-drive file=overlay.qcow2,format=qcow2,if=none,id=drive0,cache=writethrough,discard=unmap,detect-zeroes=unmap,aio=io_uring \
-device virtio-blk-device,drive=drive0,iothread=iothread0

Flag	Impact
iothread	Disk I/O runs on a dedicated thread, not QEMU's main loop. Prevents disk operations from blocking CPU emulation or display rendering.
aio=io_uring	Linux's modern async I/O API (kernel 5.8+). ~50% lower disk latency vs the default thread-pool AIO. Only on Linux; macOS/Windows use QEMU's default (threads).
virtio-blk	Paravirtualized block device — the guest knows it's virtual and uses an optimized driver. Much faster than emulated SATA/IDE.
cache=writethrough	Writes hit disk immediately. Slower than writeback but safe for snapshots — no dirty cache to lose on pivot.
discard=unmap	Guest TRIM commands release qcow2 clusters back to the host filesystem. Keeps overlays from growing indefinitely.
detect-zeroes=unmap	Writing a block of zeroes is treated as a discard. Common after fstrim or file deletion.

Snapshot performance

Checkpoints (savevm/loadvm) capture full VM state inside the qcow2 file:

• Time is proportional to RAM size (must write all of RAM to disk)
• ~1-2 seconds for 512MB-2GB

Saves (blockdev-snapshot-sync) pivot the overlay chain:

• The pivot itself is instant (just changes where writes go)
• The copy step (writing frozen overlays to the save directory) depends on overlay size
• compress=True runs qemu-img convert which recompresses. This takes 10-30 seconds depending on data

The fstrim + qemu-io flush before a save ensures the overlay is as small as possible (freed blocks are reclaimed) and fully consistent on disk.