Proxmox VM Disk Performance | VirtIO SCSI Tuning

Slow disks in a VM are often a configuration problem, not a hardware one. Proxmox exposes several knobs — the virtual controller, per-disk IO threads, cache
mode, TRIM — and the defaults are conservative. The right combination depends on the storage underneath and how much risk you accept.

This guide changes one setting at a time and measures each. Pair it with the ZFS pool guide, the
Ceph setup, and CPU and NUMA pinning for CPU-side tuning.

Map the I/O Path

A write travels from the guest filesystem through the virtual controller into QEMU, then to the host’s storage stack. Each layer can add overhead or buffer data.

architectureio-path.txt

  guest fs --> virtual controller --> QEMU (aio, cache) --> host LVM-thin / ZFS / Ceph --> disks

Guest I/O path — Guest, controller, host storage

Benchmark First

Measure the current state with a repeatable fio job inside the guest. Save the numbers; without a baseline you cannot tell whether a change helped.

bashrandom 4k mixed

fio --name=randrw --filename=/var/tmp/fio.test --size=2G --rw=randrw --rwmixread=70 \
    --bs=4k --iodepth=32 --numjobs=4 --direct=1 --runtime=60 --time_based --group_reporting

bashsequential write

fio --name=seqw --filename=/var/tmp/fio.test --size=4G --rw=write --bs=1M --direct=1 --runtime=30 --time_based

WarningRun benchmarks on a test VM or in a quiet period; they generate heavy I/O on shared storage.

Use the VirtIO SCSI single Controller

VirtIO devices are paravirtualised and far faster than emulated IDE or SATA. VirtIO SCSI single gives each disk its own controller, which is what allows a dedicated IO thread per disk.

bashset the controller (VM shut down)

qm set 101 --scsihw virtio-scsi-single

Enable IO Threads

With iothread=1 each disk’s I/O is handled by its own thread instead of sharing QEMU’s main loop. It mostly helps VMs with several busy disks or many parallel requests.

bashenable on a disk

qm set 101 --scsi0 local-lvm:vm-101-disk-0,iothread=1

iothread — Dedicated I/O thread per disk

Pick a Cache Mode You Can Defend

none (no host caching) is the safe default and works well with modern storage. writeback is faster but can lose recent writes on power failure unless you have a UPS or battery-backed cache.

bashset cache=none explicitly

qm set 101 --scsi0 local-lvm:vm-101-disk-0,iothread=1,cache=none

WarningNever use cache=unsafe for data you care about; it ignores flush requests.

Turn On Discard and SSD Emulation

With discard=on the guest’s TRIM commands free blocks on thin-provisioned storage, so deleted data stops consuming host space. ssd=1 tells the guest to treat the disk as an SSD.

bashenable both

qm set 101 --scsi0 local-lvm:vm-101-disk-0,iothread=1,discard=on,ssd=1

bashinside a Linux guest

sudo systemctl enable --now fstrim.timer
sudo fstrim -av

Understand Async I/O and Disk Format

Recent Proxmox versions default to io_uring. Raw volumes on LVM-thin or ZFS zvols carry less overhead than qcow2 files, which trade speed for snapshots and flexibility.

bashreview the resulting config

qm config 101 | grep -E 'scsihw|scsi0'

aio and formats — io_uring and raw versus qcow2

Re-Test and Record the Result

Run the identical fio jobs after each change and keep the setting only if IOPS or latency improved. Note the final configuration alongside the storage type so the next VM starts from a known-good baseline.

bashfinal config example

scsihw: virtio-scsi-single
scsi0: local-lvm:vm-101-disk-0,iothread=1,discard=on,ssd=1,cache=none,size=64G

Verify improvements — Compare against baseline

Quick Reference

virtio-scsi-single + iothread=1; cache=none unless you have power protection
discard=on + guest fstrim; measure with fio before and after

Proxmox VM Disk Performance Tuning: VirtIO SCSI, IO Threads, Cache and Discard