i/o performance

[Venkatesh Srinivas]

On Wed, Dec 10, 2014 at 2:10 PM, <nans_nans1 at yahoo.de> wrote:

> In Proxmox the storage is configured as "virtio0".
>
> I attached a part of dmesg:
>
> atapci0: <Intel PIIX3 WDMA2 controller> port
> 0xc0a0-0xc0af,0x376,0x170-0x177,0x3f6,0x1f0-0x1f7 at device 1.1 on pci0
> ata0: <ATA channel 0> on atapci0
> ata1: <ATA channel 1> on atapci0
> disk scheduler: set policy of acd0 to noop
> acd0: DVDROM <QEMU DVD-ROM/2.1.2> at ata1-master WDMA2
> uhci0: <Intel 82371SB (PIIX3) USB controller> port 0xc040-0xc05f irq 11 at
> device 1.2 on pci0
> usbus0: controller did not stop
> usbus0: <Intel 82371SB (PIIX3) USB controller> on uhci0
> pci0: <bridge> (vendor 0x8086, dev 0x7113) at device 1.3 irq 9
> vgapci0: <VGA-compatible display> mem
> 0xfebf0000-0xfebf0fff,0xfc000000-0xfdffffff at device 2.0 on pci0
> vgapci0: Boot video device
> virtio_pci0: <VirtIO PCI Balloon adapter> port 0xc060-0xc07f irq 11 at
> device 3.0 on pci0
> virtio_pci1: <VirtIO PCI Block adapter> port 0xc000-0xc03f mem
> 0xfebf1000-0xfebf1fff irq 10 at device 10.0 on pci0
> vtblk0: <VirtIO Block Adapter> on virtio_pci1
> virtio_pci1: host features: 0x71000ed4
> <EventIdx,0x10000000,NotifyOnEmpty,0x800,Topology,FlushCmd,SCSICmds,BlockSize,DiskGeometry,MaxNumSegs>
> virtio_pci1: negotiated features: 0x254
> <FlushCmd,BlockSize,DiskGeometry,MaxNumSegs>
> disk scheduler: set policy of vbd0 to noop
>

DragonFly BSD's virtio-pci driver does not (at least 'did not') support
indirect ring entries; qemu/kvm default to a 128-entry queue for each
virtio-blk device; Dfly systems were regularly bandwidth limited by filling
that queue; (at 64 KiB I/Os and fragmented physmem, you can only have 12-16
inflight IOs per disk).

DFly also appears to not be using MSI-X for the virtio-blk device; fixing
that should result in modest but real per-I/O improvements

Bug: http://bugs.dragonflybsd.org/issues/2577 is related.

HTH,
-- vs;
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.dragonflybsd.org/pipermail/users/attachments/20141210/94db40e4/attachment-0002.html>