configure | 13 ++ hw/block/Kconfig | 5 + hw/block/meson.build | 1 + hw/block/vhost-blk.c | 395 ++++++++++++++++++++++++++++++++++ hw/virtio/meson.build | 1 + hw/virtio/vhost-blk-pci.c | 102 +++++++++ include/hw/virtio/vhost-blk.h | 44 ++++ linux-headers/linux/vhost.h | 3 + 8 files changed, 564 insertions(+) create mode 100644 hw/block/vhost-blk.c create mode 100644 hw/virtio/vhost-blk-pci.c create mode 100644 include/hw/virtio/vhost-blk.h
Although QEMU virtio-blk is quite fast, there is still some room for improvements. Disk latency can be reduced if we handle virito-blk requests in host kernel so we avoid a lot of syscalls and context switches. The biggest disadvantage of this vhost-blk flavor is raw format. Luckily Kirill Thai proposed device mapper driver for QCOW2 format to attach files as block devices: https://www.spinics.net/lists/kernel/msg4292965.html Also by using kernel modules we can bypass iothread limitation and finaly scale block requests with cpus for high-performance devices. This is planned to be implemented in next version. Linux kernel module part: https://lore.kernel.org/kvm/20220725202753.298725-1-andrey.zhadchenko@virtuozzo.com/ test setups and results: fio --direct=1 --rw=randread --bs=4k --ioengine=libaio --iodepth=128 QEMU drive options: cache=none filesystem: xfs SSD: | randread, IOPS | randwrite, IOPS | Host | 95.8k | 85.3k | QEMU virtio | 57.5k | 79.4k | QEMU vhost-blk | 95.6k | 84.3k | RAMDISK (vq == vcpu): | randread, IOPS | randwrite, IOPS | virtio, 1vcpu | 123k | 129k | virtio, 2vcpu | 253k (??) | 250k (??) | virtio, 4vcpu | 158k | 154k | vhost-blk, 1vcpu | 110k | 113k | vhost-blk, 2vcpu | 247k | 252k | vhost-blk, 4vcpu | 576k | 567k | Andrey Zhadchenko (1): block: add vhost-blk backend configure | 13 ++ hw/block/Kconfig | 5 + hw/block/meson.build | 1 + hw/block/vhost-blk.c | 395 ++++++++++++++++++++++++++++++++++ hw/virtio/meson.build | 1 + hw/virtio/vhost-blk-pci.c | 102 +++++++++ include/hw/virtio/vhost-blk.h | 44 ++++ linux-headers/linux/vhost.h | 3 + 8 files changed, 564 insertions(+) create mode 100644 hw/block/vhost-blk.c create mode 100644 hw/virtio/vhost-blk-pci.c create mode 100644 include/hw/virtio/vhost-blk.h -- 2.31.1
On Mon, Jul 25, 2022 at 11:55:26PM +0300, Andrey Zhadchenko wrote: > Although QEMU virtio-blk is quite fast, there is still some room for > improvements. Disk latency can be reduced if we handle virito-blk requests > in host kernel so we avoid a lot of syscalls and context switches. > > The biggest disadvantage of this vhost-blk flavor is raw format. > Luckily Kirill Thai proposed device mapper driver for QCOW2 format to attach > files as block devices: https://www.spinics.net/lists/kernel/msg4292965.html > > Also by using kernel modules we can bypass iothread limitation and finaly scale > block requests with cpus for high-performance devices. This is planned to be > implemented in next version. > > Linux kernel module part: > https://lore.kernel.org/kvm/20220725202753.298725-1-andrey.zhadchenko@virtuozzo.com/ > > test setups and results: > fio --direct=1 --rw=randread --bs=4k --ioengine=libaio --iodepth=128 > QEMU drive options: cache=none > filesystem: xfs Please post the full QEMU command-line so it's clear exactly what this is benchmarking. A preallocated raw image file is a good baseline with: --object iothread,id=iothread0 \ --blockdev file,filename=test.img,cache.direct=on,aio=native,node-name=drive0 \ --device virtio-blk-pci,drive=drive0,iothread=iothread0 (BTW QEMU's default vq size is 256 descriptors and the number of vqs is the number of vCPUs.) > > SSD: > | randread, IOPS | randwrite, IOPS | > Host | 95.8k | 85.3k | > QEMU virtio | 57.5k | 79.4k | > QEMU vhost-blk | 95.6k | 84.3k | > > RAMDISK (vq == vcpu): With fio numjobs=vcpu here? > | randread, IOPS | randwrite, IOPS | > virtio, 1vcpu | 123k | 129k | > virtio, 2vcpu | 253k (??) | 250k (??) | QEMU's aio=threads (default) gets around the single IOThread. It beats aio=native for this reason in some cases. Were you using aio=native or aio=threads? > virtio, 4vcpu | 158k | 154k | > vhost-blk, 1vcpu | 110k | 113k | > vhost-blk, 2vcpu | 247k | 252k |
On 10/4/22 21:26, Stefan Hajnoczi wrote: > On Mon, Jul 25, 2022 at 11:55:26PM +0300, Andrey Zhadchenko wrote: >> Although QEMU virtio-blk is quite fast, there is still some room for >> improvements. Disk latency can be reduced if we handle virito-blk requests >> in host kernel so we avoid a lot of syscalls and context switches. >> >> The biggest disadvantage of this vhost-blk flavor is raw format. >> Luckily Kirill Thai proposed device mapper driver for QCOW2 format to attach >> files as block devices: https://www.spinics.net/lists/kernel/msg4292965.html >> >> Also by using kernel modules we can bypass iothread limitation and finaly scale >> block requests with cpus for high-performance devices. This is planned to be >> implemented in next version. >> >> Linux kernel module part: >> https://lore.kernel.org/kvm/20220725202753.298725-1-andrey.zhadchenko@virtuozzo.com/ >> >> test setups and results: >> fio --direct=1 --rw=randread --bs=4k --ioengine=libaio --iodepth=128 > >> QEMU drive options: cache=none >> filesystem: xfs > > Please post the full QEMU command-line so it's clear exactly what this > is benchmarking. The full command for vhost is this: qemu-system-x86_64 \ -kernel bzImage -nographic -append "console=ttyS0 root=/dev/sdb rw systemd.unified_cgroup_hierarchy=0 nokaslr" \ -m 1024 -s --enable-kvm -smp $2 \ -drive id=main_drive,file=debian_sid.img,media=disk,format=raw \ -drive id=vhost_drive,file=$1,media=disk,format=raw,if=none \ -device vhost-blk-pci,drive=vhost_drive,num-threads=$3 (num-threads option for vhost-blk-pci was not used) For virtio I used this: qemu-system-x86_64 \ -kernel bzImage -nographic -append "console=ttyS0 root=/dev/sdb rw systemd.unified_cgroup_hierarchy=0 nokaslr" \ -m 1024 -s --enable-kvm -smp $2 \ -drive file=debian_sid.img,media=disk \ -drive file=$1,media=disk,if=virtio,cache=none,if=none,id=d1,aio=threads\ -device virtio-blk-pci,drive=d1 > > A preallocated raw image file is a good baseline with: > > --object iothread,id=iothread0 \ > --blockdev file,filename=test.img,cache.direct=on,aio=native,node-name=drive0 > --device virtio-blk-pci,drive=drive0,iothread=iothread0 The image I used was preallocated qcow2 image set up with dm-qcow2 because this vhost-blk version directly uses bio interface and can't work with regular files. > > (BTW QEMU's default vq size is 256 descriptors and the number of vqs is > the number of vCPUs.) > >> >> SSD: >> | randread, IOPS | randwrite, IOPS | >> Host | 95.8k | 85.3k | >> QEMU virtio | 57.5k | 79.4k | Adding iothread0 and using raw file instead of qcow2 + dm-qcow2 setup brings the numbers to | 60.4k | 84.3k | >> QEMU vhost-blk | 95.6k | 84.3k | >> >> RAMDISK (vq == vcpu): > > With fio numjobs=vcpu here? Yes > >> | randread, IOPS | randwrite, IOPS | >> virtio, 1vcpu | 123k | 129k | >> virtio, 2vcpu | 253k (??) | 250k (??) | > > QEMU's aio=threads (default) gets around the single IOThread. It beats > aio=native for this reason in some cases. Were you using aio=native or > aio=threads? At some point of time I started to specify aio=threads (and before that I did not use this option). I am not sure when exactly. I will re-measure all cases for the next submission. > >> virtio, 4vcpu | 158k | 154k | >> vhost-blk, 1vcpu | 110k | 113k | >> vhost-blk, 2vcpu | 247k | 252k |
On Wed, Oct 05, 2022 at 01:28:14PM +0300, Andrey Zhadchenko wrote: > > > On 10/4/22 21:26, Stefan Hajnoczi wrote: > > On Mon, Jul 25, 2022 at 11:55:26PM +0300, Andrey Zhadchenko wrote: > > > Although QEMU virtio-blk is quite fast, there is still some room for > > > improvements. Disk latency can be reduced if we handle virito-blk requests > > > in host kernel so we avoid a lot of syscalls and context switches. > > > > > > The biggest disadvantage of this vhost-blk flavor is raw format. > > > Luckily Kirill Thai proposed device mapper driver for QCOW2 format to attach > > > files as block devices: https://www.spinics.net/lists/kernel/msg4292965.html > > > > > > Also by using kernel modules we can bypass iothread limitation and finaly scale > > > block requests with cpus for high-performance devices. This is planned to be > > > implemented in next version. > > > > > > Linux kernel module part: > > > https://lore.kernel.org/kvm/20220725202753.298725-1-andrey.zhadchenko@virtuozzo.com/ > > > > > > test setups and results: > > > fio --direct=1 --rw=randread --bs=4k --ioengine=libaio --iodepth=128 > > > > > QEMU drive options: cache=none > > > filesystem: xfs > > > > Please post the full QEMU command-line so it's clear exactly what this > > is benchmarking. > > The full command for vhost is this: > qemu-system-x86_64 \ > -kernel bzImage -nographic -append "console=ttyS0 root=/dev/sdb rw > systemd.unified_cgroup_hierarchy=0 nokaslr" \ > -m 1024 -s --enable-kvm -smp $2 \ > -drive id=main_drive,file=debian_sid.img,media=disk,format=raw \ > -drive id=vhost_drive,file=$1,media=disk,format=raw,if=none \ No cache=none because vhost-blk directly submits bios in the kernel? > -device vhost-blk-pci,drive=vhost_drive,num-threads=$3 > > (num-threads option for vhost-blk-pci was not used) > > For virtio I used this: > qemu-system-x86_64 \ > -kernel bzImage -nographic -append "console=ttyS0 root=/dev/sdb rw > systemd.unified_cgroup_hierarchy=0 nokaslr" \ > -m 1024 -s --enable-kvm -smp $2 \ > -drive file=debian_sid.img,media=disk \ > -drive file=$1,media=disk,if=virtio,cache=none,if=none,id=d1,aio=threads\ > -device virtio-blk-pci,drive=d1 > > > > > A preallocated raw image file is a good baseline with: > > > > --object iothread,id=iothread0 \ > > --blockdev file,filename=test.img,cache.direct=on,aio=native,node-name=drive0 > --device virtio-blk-pci,drive=drive0,iothread=iothread0 > The image I used was preallocated qcow2 image set up with dm-qcow2 because > this vhost-blk version directly uses bio interface and can't work with > regular files. I see. > > > > > (BTW QEMU's default vq size is 256 descriptors and the number of vqs is > > the number of vCPUs.) > > > > > > > > SSD: > > > | randread, IOPS | randwrite, IOPS | > > > Host | 95.8k | 85.3k | > > > QEMU virtio | 57.5k | 79.4k | > > Adding iothread0 and using raw file instead of qcow2 + dm-qcow2 setup brings > the numbers to > | 60.4k | 84.3k | > > > > QEMU vhost-blk | 95.6k | 84.3k | > > > > > > RAMDISK (vq == vcpu): > > > > With fio numjobs=vcpu here? > > Yes > > > > > > | randread, IOPS | randwrite, IOPS | > > > virtio, 1vcpu | 123k | 129k | > > > virtio, 2vcpu | 253k (??) | 250k (??) | > > > > QEMU's aio=threads (default) gets around the single IOThread. It beats > > aio=native for this reason in some cases. Were you using aio=native or > > aio=threads? > > At some point of time I started to specify aio=threads (and before that I > did not use this option). I am not sure when exactly. I will re-measure all > cases for the next submission. aio=native is usually recommended. aio=threads is less optimized. aio=native should have lower latency than aio=threads although it scales worse on hosts with free CPUs because it's limited to a single thread. > > > > > > virtio, 4vcpu | 158k | 154k | > > > vhost-blk, 1vcpu | 110k | 113k | > > > vhost-blk, 2vcpu | 247k | 252k | >
On Mon, Jul 25, 2022 at 11:55:26PM +0300, Andrey Zhadchenko wrote: > Although QEMU virtio-blk is quite fast, there is still some room for > improvements. Disk latency can be reduced if we handle virito-blk requests > in host kernel so we avoid a lot of syscalls and context switches. > > The biggest disadvantage of this vhost-blk flavor is raw format. > Luckily Kirill Thai proposed device mapper driver for QCOW2 format to attach > files as block devices: https://www.spinics.net/lists/kernel/msg4292965.html > > Also by using kernel modules we can bypass iothread limitation and finaly scale > block requests with cpus for high-performance devices. This is planned to be > implemented in next version. > > Linux kernel module part: > https://lore.kernel.org/kvm/20220725202753.298725-1-andrey.zhadchenko@virtuozzo.com/ > > test setups and results: > fio --direct=1 --rw=randread --bs=4k --ioengine=libaio --iodepth=128 > QEMU drive options: cache=none > filesystem: xfs > > SSD: > | randread, IOPS | randwrite, IOPS | > Host | 95.8k | 85.3k | > QEMU virtio | 57.5k | 79.4k | > QEMU vhost-blk | 95.6k | 84.3k | > > RAMDISK (vq == vcpu): > | randread, IOPS | randwrite, IOPS | > virtio, 1vcpu | 123k | 129k | > virtio, 2vcpu | 253k (??) | 250k (??) | > virtio, 4vcpu | 158k | 154k | > vhost-blk, 1vcpu | 110k | 113k | > vhost-blk, 2vcpu | 247k | 252k | > vhost-blk, 4vcpu | 576k | 567k | > > Andrey Zhadchenko (1): > block: add vhost-blk backend > > configure | 13 ++ > hw/block/Kconfig | 5 + > hw/block/meson.build | 1 + > hw/block/vhost-blk.c | 395 ++++++++++++++++++++++++++++++++++ > hw/virtio/meson.build | 1 + > hw/virtio/vhost-blk-pci.c | 102 +++++++++ > include/hw/virtio/vhost-blk.h | 44 ++++ > linux-headers/linux/vhost.h | 3 + > 8 files changed, 564 insertions(+) > create mode 100644 hw/block/vhost-blk.c > create mode 100644 hw/virtio/vhost-blk-pci.c > create mode 100644 include/hw/virtio/vhost-blk.h vhost-blk has been tried several times in the past. That doesn't mean it cannot be merged this time, but past arguments should be addressed: - What makes it necessary to move the code into the kernel? In the past the performance results were not very convincing. The fastest implementations actually tend to be userspace NVMe PCI drivers that bypass the kernel! Bypassing the VFS and submitting block requests directly was not a huge boost. The syscall/context switch argument sounds okay but the numbers didn't really show that kernel block I/O is much faster than userspace block I/O. I've asked for more details on the QEMU command-line to understand what your numbers show. Maybe something has changed since previous times when vhost-blk has been tried. The only argument I see is QEMU's current 1 IOThread per virtio-blk device limitation, which is currently being worked on. If that's the only reason for vhost-blk then is it worth doing all the work of getting vhost-blk shipped (kernel, QEMU, and libvirt changes)? It seems like a short-term solution. - The security impact of bugs in kernel vhost-blk code is more serious than bugs in a QEMU userspace process. - The management stack needs to be changed to use vhost-blk whereas QEMU can be optimized without affecting other layers. Stefan
On 10/4/22 22:00, Stefan Hajnoczi wrote: > On Mon, Jul 25, 2022 at 11:55:26PM +0300, Andrey Zhadchenko wrote: >> Although QEMU virtio-blk is quite fast, there is still some room for >> improvements. Disk latency can be reduced if we handle virito-blk requests >> in host kernel so we avoid a lot of syscalls and context switches. >> >> The biggest disadvantage of this vhost-blk flavor is raw format. >> Luckily Kirill Thai proposed device mapper driver for QCOW2 format to attach >> files as block devices: https://www.spinics.net/lists/kernel/msg4292965.html >> >> Also by using kernel modules we can bypass iothread limitation and finaly scale >> block requests with cpus for high-performance devices. This is planned to be >> implemented in next version. >> >> Linux kernel module part: >> https://lore.kernel.org/kvm/20220725202753.298725-1-andrey.zhadchenko@virtuozzo.com/ >> >> test setups and results: >> fio --direct=1 --rw=randread --bs=4k --ioengine=libaio --iodepth=128 >> QEMU drive options: cache=none >> filesystem: xfs >> >> SSD: >> | randread, IOPS | randwrite, IOPS | >> Host | 95.8k | 85.3k | >> QEMU virtio | 57.5k | 79.4k | >> QEMU vhost-blk | 95.6k | 84.3k | >> >> RAMDISK (vq == vcpu): >> | randread, IOPS | randwrite, IOPS | >> virtio, 1vcpu | 123k | 129k | >> virtio, 2vcpu | 253k (??) | 250k (??) | >> virtio, 4vcpu | 158k | 154k | >> vhost-blk, 1vcpu | 110k | 113k | >> vhost-blk, 2vcpu | 247k | 252k | >> vhost-blk, 4vcpu | 576k | 567k | >> >> Andrey Zhadchenko (1): >> block: add vhost-blk backend >> >> configure | 13 ++ >> hw/block/Kconfig | 5 + >> hw/block/meson.build | 1 + >> hw/block/vhost-blk.c | 395 ++++++++++++++++++++++++++++++++++ >> hw/virtio/meson.build | 1 + >> hw/virtio/vhost-blk-pci.c | 102 +++++++++ >> include/hw/virtio/vhost-blk.h | 44 ++++ >> linux-headers/linux/vhost.h | 3 + >> 8 files changed, 564 insertions(+) >> create mode 100644 hw/block/vhost-blk.c >> create mode 100644 hw/virtio/vhost-blk-pci.c >> create mode 100644 include/hw/virtio/vhost-blk.h > > vhost-blk has been tried several times in the past. That doesn't mean it > cannot be merged this time, but past arguments should be addressed: > > - What makes it necessary to move the code into the kernel? In the past > the performance results were not very convincing. The fastest > implementations actually tend to be userspace NVMe PCI drivers that > bypass the kernel! Bypassing the VFS and submitting block requests > directly was not a huge boost. The syscall/context switch argument > sounds okay but the numbers didn't really show that kernel block I/O > is much faster than userspace block I/O. > > I've asked for more details on the QEMU command-line to understand > what your numbers show. Maybe something has changed since previous > times when vhost-blk has been tried. > > The only argument I see is QEMU's current 1 IOThread per virtio-blk > device limitation, which is currently being worked on. If that's the > only reason for vhost-blk then is it worth doing all the work of > getting vhost-blk shipped (kernel, QEMU, and libvirt changes)? It > seems like a short-term solution. > > - The security impact of bugs in kernel vhost-blk code is more serious > than bugs in a QEMU userspace process. > > - The management stack needs to be changed to use vhost-blk whereas > QEMU can be optimized without affecting other layers. > > Stefan Indeed there was several vhost-blk attempts, but from what I found in mailing lists only the Asias attempt got some attention and discussion. Ramdisk performance results were great but ramdisk is more a benchmark than a real use. I didn't find out why Asias dropped his version except vague "He concluded performance results was not worth". The storage speed is very important for vhost-blk performance, as there is no point to cut cpu costs from 1ms to 0,1ms if the request need 50ms to proceed in the actual disk. I think that 10 years ago NVMI was non-existent and SSD + SATA was probably a lot faster than HDD but still not enough to utilize this technology. The tests I did give me 60k IOPS randwrite for VM and 95k for host. And the vhost-blk is able to negate the difference even using only 1 thread/vq/vcpu. And unlinke current QEMU single IOThread it can be easily scaled with number of cpus/vcpus. For sure this can be solved by liftimg IOThread limitations but this will probably be even more disastrous amount of changes (and adding vhost-blk won't break old setups!). Probably the only undisputed advantage of vhost-blk is syscalls reduction. And again the benefit really depends on a storage speed, as it should be somehow comparable with syscalls time. Also I must note that this may be good for high-density servers with a lot of VMs. But for now I did not have the exact numbers which show how much time we are really winning for a single request at average. Overall vhost-blk will only become better along with the increase of storage speed. Also I must note that all arguments above apply to vdpa-blk. And unlike vhost-blk, which needs it's own QEMU code, vdpa-blk can be setup with generic virtio-vdpa QEMU code (I am not sure if it is merged yet but still). Although vdpa-blk have it's own problems for now.
On Wed, Oct 05, 2022 at 02:50:06PM +0300, Andrey Zhadchenko wrote: > > > On 10/4/22 22:00, Stefan Hajnoczi wrote: > > On Mon, Jul 25, 2022 at 11:55:26PM +0300, Andrey Zhadchenko wrote: > > > Although QEMU virtio-blk is quite fast, there is still some room for > > > improvements. Disk latency can be reduced if we handle virito-blk requests > > > in host kernel so we avoid a lot of syscalls and context switches. > > > > > > The biggest disadvantage of this vhost-blk flavor is raw format. > > > Luckily Kirill Thai proposed device mapper driver for QCOW2 format to attach > > > files as block devices: https://www.spinics.net/lists/kernel/msg4292965.html > > > > > > Also by using kernel modules we can bypass iothread limitation and finaly scale > > > block requests with cpus for high-performance devices. This is planned to be > > > implemented in next version. > > > > > > Linux kernel module part: > > > https://lore.kernel.org/kvm/20220725202753.298725-1-andrey.zhadchenko@virtuozzo.com/ > > > > > > test setups and results: > > > fio --direct=1 --rw=randread --bs=4k --ioengine=libaio --iodepth=128 > > > QEMU drive options: cache=none > > > filesystem: xfs > > > > > > SSD: > > > | randread, IOPS | randwrite, IOPS | > > > Host | 95.8k | 85.3k | > > > QEMU virtio | 57.5k | 79.4k | > > > QEMU vhost-blk | 95.6k | 84.3k | > > > > > > RAMDISK (vq == vcpu): > > > | randread, IOPS | randwrite, IOPS | > > > virtio, 1vcpu | 123k | 129k | > > > virtio, 2vcpu | 253k (??) | 250k (??) | > > > virtio, 4vcpu | 158k | 154k | > > > vhost-blk, 1vcpu | 110k | 113k | > > > vhost-blk, 2vcpu | 247k | 252k | > > > vhost-blk, 4vcpu | 576k | 567k | > > > > > > Andrey Zhadchenko (1): > > > block: add vhost-blk backend > > > > > > configure | 13 ++ > > > hw/block/Kconfig | 5 + > > > hw/block/meson.build | 1 + > > > hw/block/vhost-blk.c | 395 ++++++++++++++++++++++++++++++++++ > > > hw/virtio/meson.build | 1 + > > > hw/virtio/vhost-blk-pci.c | 102 +++++++++ > > > include/hw/virtio/vhost-blk.h | 44 ++++ > > > linux-headers/linux/vhost.h | 3 + > > > 8 files changed, 564 insertions(+) > > > create mode 100644 hw/block/vhost-blk.c > > > create mode 100644 hw/virtio/vhost-blk-pci.c > > > create mode 100644 include/hw/virtio/vhost-blk.h > > > > vhost-blk has been tried several times in the past. That doesn't mean it > > cannot be merged this time, but past arguments should be addressed: > > > > - What makes it necessary to move the code into the kernel? In the past > > the performance results were not very convincing. The fastest > > implementations actually tend to be userspace NVMe PCI drivers that > > bypass the kernel! Bypassing the VFS and submitting block requests > > directly was not a huge boost. The syscall/context switch argument > > sounds okay but the numbers didn't really show that kernel block I/O > > is much faster than userspace block I/O. > > > > I've asked for more details on the QEMU command-line to understand > > what your numbers show. Maybe something has changed since previous > > times when vhost-blk has been tried. > > > > The only argument I see is QEMU's current 1 IOThread per virtio-blk > > device limitation, which is currently being worked on. If that's the > > only reason for vhost-blk then is it worth doing all the work of > > getting vhost-blk shipped (kernel, QEMU, and libvirt changes)? It > > seems like a short-term solution. > > > > - The security impact of bugs in kernel vhost-blk code is more serious > > than bugs in a QEMU userspace process. > > > > - The management stack needs to be changed to use vhost-blk whereas > > QEMU can be optimized without affecting other layers. > > > > Stefan > > Indeed there was several vhost-blk attempts, but from what I found in > mailing lists only the Asias attempt got some attention and discussion. > Ramdisk performance results were great but ramdisk is more a benchmark than > a real use. I didn't find out why Asias dropped his version except vague "He > concluded performance results was not worth". The storage speed is very > important for vhost-blk performance, as there is no point to cut cpu costs > from 1ms to 0,1ms if the request need 50ms to proceed in the actual disk. I > think that 10 years ago NVMI was non-existent and SSD + SATA was probably a > lot faster than HDD but still not enough to utilize this technology. Yes, it's possible that latency improvements are more noticeable now. Thank you for posting the benchmark results. I will also run benchmarks so we can compare vhost-blk with today's QEMU as well as multiqueue IOThreads QEMU (for which I only have a hacky prototype) on a local NVMe PCI SSD. > The tests I did give me 60k IOPS randwrite for VM and 95k for host. And the > vhost-blk is able to negate the difference even using only 1 thread/vq/vcpu. > And unlinke current QEMU single IOThread it can be easily scaled with number > of cpus/vcpus. For sure this can be solved by liftimg IOThread limitations > but this will probably be even more disastrous amount of changes (and adding > vhost-blk won't break old setups!). > > Probably the only undisputed advantage of vhost-blk is syscalls reduction. > And again the benefit really depends on a storage speed, as it should be > somehow comparable with syscalls time. Also I must note that this may be > good for high-density servers with a lot of VMs. But for now I did not have > the exact numbers which show how much time we are really winning for a > single request at average. > > Overall vhost-blk will only become better along with the increase of storage > speed. > > Also I must note that all arguments above apply to vdpa-blk. And unlike > vhost-blk, which needs it's own QEMU code, vdpa-blk can be setup with > generic virtio-vdpa QEMU code (I am not sure if it is merged yet but still). > Although vdpa-blk have it's own problems for now. Yes, I think that's why Stefano hasn't pushed for a software vpda-blk device yet despite having played with it and is more focussed on hardware enablement. vdpa-blk has the same issues as vhost-blk. Stefan
On Mon, Jul 25, 2022 at 11:55:26PM +0300, Andrey Zhadchenko wrote: > Although QEMU virtio-blk is quite fast, there is still some room for > improvements. Disk latency can be reduced if we handle virito-blk requests > in host kernel so we avoid a lot of syscalls and context switches. > > The biggest disadvantage of this vhost-blk flavor is raw format. > Luckily Kirill Thai proposed device mapper driver for QCOW2 format to attach > files as block devices: https://www.spinics.net/lists/kernel/msg4292965.html > > Also by using kernel modules we can bypass iothread limitation and finaly scale > block requests with cpus for high-performance devices. This is planned to be > implemented in next version. Hi Andrey, Do you have a new version of this patch series that uses multiple threads? I have been playing with vq-IOThread mapping in QEMU and would like to benchmark vhost-blk vs QEMU virtio-blk mq IOThreads: https://gitlab.com/stefanha/qemu/-/tree/virtio-blk-mq-iothread-prototype Thanks, Stefan
On 10/4/22 21:13, Stefan Hajnoczi wrote: > On Mon, Jul 25, 2022 at 11:55:26PM +0300, Andrey Zhadchenko wrote: >> Although QEMU virtio-blk is quite fast, there is still some room for >> improvements. Disk latency can be reduced if we handle virito-blk requests >> in host kernel so we avoid a lot of syscalls and context switches. >> >> The biggest disadvantage of this vhost-blk flavor is raw format. >> Luckily Kirill Thai proposed device mapper driver for QCOW2 format to attach >> files as block devices: https://www.spinics.net/lists/kernel/msg4292965.html >> >> Also by using kernel modules we can bypass iothread limitation and finaly scale >> block requests with cpus for high-performance devices. This is planned to be >> implemented in next version. > > Hi Andrey, > Do you have a new version of this patch series that uses multiple > threads? > > I have been playing with vq-IOThread mapping in QEMU and would like to > benchmark vhost-blk vs QEMU virtio-blk mq IOThreads: > https://gitlab.com/stefanha/qemu/-/tree/virtio-blk-mq-iothread-prototype > > Thanks, > Stefan Hi Stefan For now my multi-threaded version is only available for Red Hat 9 5.14.0 kernel. If you really want you can grab it from here: https://lists.openvz.org/pipermail/devel/2022-September/079951.html (kernel) For QEMU part all you need is adding to vhost_blk_start something like: #define VHOST_SET_NWORKERS _IOW(VHOST_VIRTIO, 0x1F, int) ioctl(s->vhostfd, VHOST_SET_NWORKERS, &nworkers); Or you can wait a bit. I should be able to send second versions by the end of the week (Monday in worst case). Thanks, Andrey
On Wed, Oct 05, 2022 at 12:14:18PM +0300, Andrey Zhadchenko wrote: > On 10/4/22 21:13, Stefan Hajnoczi wrote: > > On Mon, Jul 25, 2022 at 11:55:26PM +0300, Andrey Zhadchenko wrote: > > > Although QEMU virtio-blk is quite fast, there is still some room for > > > improvements. Disk latency can be reduced if we handle virito-blk requests > > > in host kernel so we avoid a lot of syscalls and context switches. > > > > > > The biggest disadvantage of this vhost-blk flavor is raw format. > > > Luckily Kirill Thai proposed device mapper driver for QCOW2 format to attach > > > files as block devices: https://www.spinics.net/lists/kernel/msg4292965.html > > > > > > Also by using kernel modules we can bypass iothread limitation and finaly scale > > > block requests with cpus for high-performance devices. This is planned to be > > > implemented in next version. > > > > Hi Andrey, > > Do you have a new version of this patch series that uses multiple > > threads? > > > > I have been playing with vq-IOThread mapping in QEMU and would like to > > benchmark vhost-blk vs QEMU virtio-blk mq IOThreads: > > https://gitlab.com/stefanha/qemu/-/tree/virtio-blk-mq-iothread-prototype > > > > Thanks, > > Stefan > > Hi Stefan > For now my multi-threaded version is only available for Red Hat 9 5.14.0 > kernel. If you really want you can grab it from here: > https://lists.openvz.org/pipermail/devel/2022-September/079951.html (kernel) > For QEMU part all you need is adding to vhost_blk_start something like: > > #define VHOST_SET_NWORKERS _IOW(VHOST_VIRTIO, 0x1F, int) > ioctl(s->vhostfd, VHOST_SET_NWORKERS, &nworkers); > > Or you can wait a bit. I should be able to send second versions by the end > of the week (Monday in worst case). Thanks, I will wait for the next revision. Stefan
On Mon, Jul 25, 2022 at 11:55:26PM +0300, Andrey Zhadchenko wrote: > Although QEMU virtio-blk is quite fast, there is still some room for > improvements. Disk latency can be reduced if we handle virito-blk requests > in host kernel so we avoid a lot of syscalls and context switches. > > The biggest disadvantage of this vhost-blk flavor is raw format. > Luckily Kirill Thai proposed device mapper driver for QCOW2 format to attach > files as block devices: https://www.spinics.net/lists/kernel/msg4292965.html That one seems stalled. Do you plan to work on that too? > Also by using kernel modules we can bypass iothread limitation and finaly scale > block requests with cpus for high-performance devices. This is planned to be > implemented in next version. > > Linux kernel module part: > https://lore.kernel.org/kvm/20220725202753.298725-1-andrey.zhadchenko@virtuozzo.com/ > > test setups and results: > fio --direct=1 --rw=randread --bs=4k --ioengine=libaio --iodepth=128 > QEMU drive options: cache=none > filesystem: xfs > > SSD: > | randread, IOPS | randwrite, IOPS | > Host | 95.8k | 85.3k | > QEMU virtio | 57.5k | 79.4k | > QEMU vhost-blk | 95.6k | 84.3k | > > RAMDISK (vq == vcpu): > | randread, IOPS | randwrite, IOPS | > virtio, 1vcpu | 123k | 129k | > virtio, 2vcpu | 253k (??) | 250k (??) | > virtio, 4vcpu | 158k | 154k | > vhost-blk, 1vcpu | 110k | 113k | > vhost-blk, 2vcpu | 247k | 252k | > vhost-blk, 4vcpu | 576k | 567k | > > Andrey Zhadchenko (1): > block: add vhost-blk backend From vhost/virtio side the patchset looks ok. But let's see what do block devs think about it. > configure | 13 ++ > hw/block/Kconfig | 5 + > hw/block/meson.build | 1 + > hw/block/vhost-blk.c | 395 ++++++++++++++++++++++++++++++++++ > hw/virtio/meson.build | 1 + > hw/virtio/vhost-blk-pci.c | 102 +++++++++ > include/hw/virtio/vhost-blk.h | 44 ++++ > linux-headers/linux/vhost.h | 3 + > 8 files changed, 564 insertions(+) > create mode 100644 hw/block/vhost-blk.c > create mode 100644 hw/virtio/vhost-blk-pci.c > create mode 100644 include/hw/virtio/vhost-blk.h > > -- > 2.31.1
On 26.07.2022 15:51, Michael S. Tsirkin wrote: > On Mon, Jul 25, 2022 at 11:55:26PM +0300, Andrey Zhadchenko wrote: >> Although QEMU virtio-blk is quite fast, there is still some room for >> improvements. Disk latency can be reduced if we handle virito-blk requests >> in host kernel so we avoid a lot of syscalls and context switches. >> >> The biggest disadvantage of this vhost-blk flavor is raw format. >> Luckily Kirill Thai proposed device mapper driver for QCOW2 format to attach >> files as block devices: https://www.spinics.net/lists/kernel/msg4292965.html > That one seems stalled. Do you plan to work on that too? We have too. The difference in numbers, as you seen below is quite too much. We have waited for this patch to be sent to keep pushing. It should be noted that may be talk on OSS this year could also push a bit. Den >> Also by using kernel modules we can bypass iothread limitation and finaly scale >> block requests with cpus for high-performance devices. This is planned to be >> implemented in next version. >> >> Linux kernel module part: >> https://lore.kernel.org/kvm/20220725202753.298725-1-andrey.zhadchenko@virtuozzo.com/ >> >> test setups and results: >> fio --direct=1 --rw=randread --bs=4k --ioengine=libaio --iodepth=128 >> QEMU drive options: cache=none >> filesystem: xfs >> >> SSD: >> | randread, IOPS | randwrite, IOPS | >> Host | 95.8k | 85.3k | >> QEMU virtio | 57.5k | 79.4k | >> QEMU vhost-blk | 95.6k | 84.3k | >> >> RAMDISK (vq == vcpu): >> | randread, IOPS | randwrite, IOPS | >> virtio, 1vcpu | 123k | 129k | >> virtio, 2vcpu | 253k (??) | 250k (??) | >> virtio, 4vcpu | 158k | 154k | >> vhost-blk, 1vcpu | 110k | 113k | >> vhost-blk, 2vcpu | 247k | 252k | >> vhost-blk, 4vcpu | 576k | 567k | >> >> Andrey Zhadchenko (1): >> block: add vhost-blk backend > > From vhost/virtio side the patchset looks ok. But let's see what do > block devs think about it. > > >> configure | 13 ++ >> hw/block/Kconfig | 5 + >> hw/block/meson.build | 1 + >> hw/block/vhost-blk.c | 395 ++++++++++++++++++++++++++++++++++ >> hw/virtio/meson.build | 1 + >> hw/virtio/vhost-blk-pci.c | 102 +++++++++ >> include/hw/virtio/vhost-blk.h | 44 ++++ >> linux-headers/linux/vhost.h | 3 + >> 8 files changed, 564 insertions(+) >> create mode 100644 hw/block/vhost-blk.c >> create mode 100644 hw/virtio/vhost-blk-pci.c >> create mode 100644 include/hw/virtio/vhost-blk.h >> >> -- >> 2.31.1
On Tue, Jul 26, 2022 at 04:15:48PM +0200, Denis V. Lunev wrote: >On 26.07.2022 15:51, Michael S. Tsirkin wrote: >>On Mon, Jul 25, 2022 at 11:55:26PM +0300, Andrey Zhadchenko wrote: >>>Although QEMU virtio-blk is quite fast, there is still some room for >>>improvements. Disk latency can be reduced if we handle virito-blk requests >>>in host kernel so we avoid a lot of syscalls and context switches. >>> >>>The biggest disadvantage of this vhost-blk flavor is raw format. >>>Luckily Kirill Thai proposed device mapper driver for QCOW2 format to attach >>>files as block devices: https://www.spinics.net/lists/kernel/msg4292965.html >>That one seems stalled. Do you plan to work on that too? >We have too. The difference in numbers, as you seen below is quite too >much. We have waited for this patch to be sent to keep pushing. > >It should be noted that may be talk on OSS this year could also push a bit. Cool, the results are similar of what I saw when I compared vhost-blk and io_uring passthrough with NVMe (Slide 7 here: [1]). About QEMU block layer support, we recently started to work on libblkio [2]. Stefan also sent an RFC [3] to implement the QEMU BlockDriver. Currently it supports virtio-blk devices using vhost-vdpa and vhost-user. We could add support for vhost (kernel) as well, though, we were thinking of leveraging vDPA to implement in-kernel software device as well. That way we could reuse a lot of the code to support both hardware and software accelerators. In the talk [1] I describe the idea a little bit, and a few months ago I did a PoC (unsubmitted RFC) to see if it was feasible and the numbers were in line with vhost-blk. Do you think we could join forces and just have an in-kernel vdpa-blk software device? Of course we could have both vhost-blk and vdpa-blk, but with vDPA perhaps we can have one software stack to maintain for both HW and software accelerators. Thanks, Stefano [1] https://kvmforum2021.sched.com/event/ke3a/vdpa-blk-unified-hardware-and-software-offload-for-virtio-blk-stefano-garzarella-red-hat [2] https://gitlab.com/libblkio/libblkio [3] https://lore.kernel.org/qemu-devel/20220708041737.1768521-1-stefanha@redhat.com/
On 7/27/22 16:06, Stefano Garzarella wrote: > On Tue, Jul 26, 2022 at 04:15:48PM +0200, Denis V. Lunev wrote: >> On 26.07.2022 15:51, Michael S. Tsirkin wrote: >>> On Mon, Jul 25, 2022 at 11:55:26PM +0300, Andrey Zhadchenko wrote: >>>> Although QEMU virtio-blk is quite fast, there is still some room for >>>> improvements. Disk latency can be reduced if we handle virito-blk >>>> requests >>>> in host kernel so we avoid a lot of syscalls and context switches. >>>> >>>> The biggest disadvantage of this vhost-blk flavor is raw format. >>>> Luckily Kirill Thai proposed device mapper driver for QCOW2 format >>>> to attach >>>> files as block devices: >>>> https://www.spinics.net/lists/kernel/msg4292965.html >>> That one seems stalled. Do you plan to work on that too? >> We have too. The difference in numbers, as you seen below is quite too >> much. We have waited for this patch to be sent to keep pushing. >> >> It should be noted that may be talk on OSS this year could also push a >> bit. > > Cool, the results are similar of what I saw when I compared vhost-blk > and io_uring passthrough with NVMe (Slide 7 here: [1]). > > About QEMU block layer support, we recently started to work on libblkio > [2]. Stefan also sent an RFC [3] to implement the QEMU BlockDriver. > Currently it supports virtio-blk devices using vhost-vdpa and vhost-user. > We could add support for vhost (kernel) as well, though, we were > thinking of leveraging vDPA to implement in-kernel software device as well. > > That way we could reuse a lot of the code to support both hardware and > software accelerators. > > In the talk [1] I describe the idea a little bit, and a few months ago I > did a PoC (unsubmitted RFC) to see if it was feasible and the numbers > were in line with vhost-blk. > > Do you think we could join forces and just have an in-kernel vdpa-blk > software device? This seems worth trying. Why double the efforts to do the same. Yet I would like to play a bit with your vdpa-blk PoC beforehand. Can you send it to me with some instructions how to run it? > > Of course we could have both vhost-blk and vdpa-blk, but with vDPA > perhaps we can have one software stack to maintain for both HW and > software accelerators. > > Thanks, > Stefano > > [1] > https://kvmforum2021.sched.com/event/ke3a/vdpa-blk-unified-hardware-and-software-offload-for-virtio-blk-stefano-garzarella-red-hat > > [2] https://gitlab.com/libblkio/libblkio > [3] > https://lore.kernel.org/qemu-devel/20220708041737.1768521-1-stefanha@redhat.com/ > >
On Thu, Jul 28, 2022 at 7:28 AM Andrey Zhadchenko <andrey.zhadchenko@virtuozzo.com> wrote: > On 7/27/22 16:06, Stefano Garzarella wrote: > > On Tue, Jul 26, 2022 at 04:15:48PM +0200, Denis V. Lunev wrote: > >> On 26.07.2022 15:51, Michael S. Tsirkin wrote: > >>> On Mon, Jul 25, 2022 at 11:55:26PM +0300, Andrey Zhadchenko wrote: > >>>> Although QEMU virtio-blk is quite fast, there is still some room for > >>>> improvements. Disk latency can be reduced if we handle virito-blk > >>>> requests > >>>> in host kernel so we avoid a lot of syscalls and context switches. > >>>> > >>>> The biggest disadvantage of this vhost-blk flavor is raw format. > >>>> Luckily Kirill Thai proposed device mapper driver for QCOW2 format > >>>> to attach > >>>> files as block devices: > >>>> https://www.spinics.net/lists/kernel/msg4292965.html > >>> That one seems stalled. Do you plan to work on that too? > >> We have too. The difference in numbers, as you seen below is quite too > >> much. We have waited for this patch to be sent to keep pushing. > >> > >> It should be noted that may be talk on OSS this year could also push a > >> bit. > > > > Cool, the results are similar of what I saw when I compared vhost-blk > > and io_uring passthrough with NVMe (Slide 7 here: [1]). > > > > About QEMU block layer support, we recently started to work on libblkio > > [2]. Stefan also sent an RFC [3] to implement the QEMU BlockDriver. > > Currently it supports virtio-blk devices using vhost-vdpa and vhost-user. > > We could add support for vhost (kernel) as well, though, we were > > thinking of leveraging vDPA to implement in-kernel software device as well. > > > > That way we could reuse a lot of the code to support both hardware and > > software accelerators. > > > > In the talk [1] I describe the idea a little bit, and a few months ago I > > did a PoC (unsubmitted RFC) to see if it was feasible and the numbers > > were in line with vhost-blk. > > > > Do you think we could join forces and just have an in-kernel vdpa-blk > > software device? > > This seems worth trying. Why double the efforts to do the same. Yet I > would like to play a bit with your vdpa-blk PoC beforehand. Great :-) > Can you send it to me with some instructions how to run it? Yep, sure! The PoC is available here: https://gitlab.com/sgarzarella/linux/-/tree/vdpa-sw-blk-poc The tree was based on Linux v5.16, but I had some issues to rebuild with new gcc, so I rebased on v5.16.20 (not tested), configs needed: CONFIG_VDPA_SW_BLOCK=m + CONFIG_VHOST_VDPA=m + dependencies. It contains: - patches required for QEMU generic vhost-vdpa support - patches to support blk_mq_ops->poll() (to use io_uring iopoll) in the guest virtio-blk driver (I used the same kernel on guest and host) - some improvements for vringh (not completed, it could be a bottleneck) - vdpa-sw and vdpa-sw-blk patches (and hacks) It is based on the vDPA simulator framework already merged upstream. The idea is to generalize the simulator to share the code between both software devices and simulators. The code needs a lot of work, I was focusing just on a working virtio-blk device emulation, but more focus on the generic part should be done. In the code there are a couple of defines to control polling. About the vdpa-blk device, you need iproute2's vdpa tool available upstream: https://wiki.linuxfoundation.org/networking/iproute2 Once the device is instantiated (see instructions later), the backend (raw file or device) can be set through a device attribute (not robust, but it was a PoC): /sys/bus/vdpa/devices/$dev_name/backend_fd I wrote a simple python script available here: https://github.com/stefano-garzarella/vm-build/blob/main/vm-tools/vdpa_set_backend_fd.py For QEMU, we are working on libblkio to support both slow path (when QEMU block layer is needed) and fast path (vqs passed directly to the device). For now libblkio supports only slow path, so to test the fast path you can use Longpeng's patches (not yet merged upstream) with generic vhost-vdpa support: https://lore.kernel.org/qemu-devel/20220514041107.1980-1-longpeng2@huawei.com/ Steps: # load vDPA block in-kernel sw device module modprobe vdpa_sw_blk # load nvme module with poll_queues set if you want to use iopoll modprobe nvme poll_queues=15 # instantiate a new vdpa-blk device vdpa dev add mgmtdev vdpasw_blk name blk0 # set backend (/dev/nvme0n1) vdpa_set_backend_fd.py -b /dev/nvme0n1 blk0 # load vhost vDPA bus ... modprobe vhost_vdpa # ... and vhost-vdpa device will appear ls -l /dev/vhost-vdpa-0 crw-------. 1 root root 510, 0 Jul 28 17:06 /dev/vhost-vdpa-0 # start QEMU patched with generic vhost-vdpa qemu-system-x86_64 ... \ -device vhost-vdpa-device-pci,vhostdev=/dev/vhost-vdpa-0 I haven't tested it recently, so I'm not sure it all works, but in the next few days I'll try. For anything else, feel free to reach me here or on IRC (sgarzare on #qemu). Thanks, Stefano
© 2016 - 2024 Red Hat, Inc.