[Qemu-discuss] Expected i/o performance for 9P shares

[Alastair Hughes]

I'm trying to automate building packages on other architectures
without needing root; as such, 9P fs shares seemed to be the best
option. However the io performance is terrible, leading to long build
times, even for guests using KVM. Is this expected? What can I do to
improve it?

I'm testing using a 100M file. The filesystem being shared is in a
tmpfs, so I get around 1GB/s for reading/writing to the file in the
host. In the guest, I get around 10-20MB/s; sometimes as low as 3MB/s.

I am currently running qemu using:
 qemu-system-i386 -enable-kvm \
 -kernel "/tmp/builder-toshiba//fs/boot/vmlinuz" \
 -append "console=ttyS0 init=/usr/bin/sinit panic=1 rootfstype=9p rw
rootflags=trans=virtio,version=9p2000.L" \
 -fsdev local,id=root,security_model=none,path=/tmp/builder-toshiba/fs \
 -device virtio-9p-pci,fsdev=root,mount_tag=/dev/root \
 -m 24 \
 -no-reboot -nographic

I tested with a raw root fs instead, and was getting around 500MB/s
r/w for a 100MB test file.
Mounting the fs as readonly gave a small performance boost
(consistently around 30MB/s).

Alastair Hughes