[pve-devel] scsihw / hotplug integration into GUI

ronnie sahlberg ronniesahlberg at gmail.com
Wed Aug 15 10:38:05 CEST 2012


On Wed, Aug 15, 2012 at 6:15 PM, Stefan Priebe - Profihost AG
<s.priebe at profihost.ag> wrote:
> Hello list,
>
>   i've cc'd ronnie as he is also interested in a benchmark regarding
> open-iscsi and libiscsi using scsi-generic.
>
> Test Environment:
> - Host and Guest Kernel 3.5.0
> - qemu-kvm git (1.2) with block and scsi-abort patch
> - libiscsi 1.5.0
> - virtio-scsi-pci
> - scsi-generic
>
>> libiscsi:
>
> rand 4k:
>   write: io=10580MB, bw=120079KB/s, iops=30019, runt= 90220msec
>   read : io=21741MB, bw=247358KB/s, iops=61839, runt= 90001msec
>
> seq 4m:
>   write: io=64828MB, bw=735678KB/s, iops=179, runt= 90235msec
>   read : io=30680MB, bw=347345KB/s, iops=84, runt= 90447msec
>
>> open-iscsi:
>
> rand 4k:
>   write: io=5814MB, bw=66096KB/s, iops=16524, runt= 90074msec
>   read : io=7524MB, bw=85602KB/s, iops=21400, runt= 90009msec
> seq 4m:
>   write: io=78460MB, bw=890494KB/s, iops=217, runt= 90223msec
>   read : io=54388MB, bw=616649KB/s, iops=150, runt= 90316msec
>
> I really want to start integrating this into Proxmox. Dietmar what's your
> opinion? Should i add another autodetection?
>
> @Ronnie
> Do you have any idea why the sequential speed of 4M blocks is lower than
> with open-iscsi?


Impressive numbers!

My best guess is that libiscsi has a little better latency than
open-iscsi and that is what shows up in the small i/o numbers.


I imagine that open-iscsi would use zero-copy for both read and write
so that should show up for big i/o sizes.
libiscsi just reads/writes from the socket() so there is at least one
data-copy when libiscsi is used.


regards
ronnie sahlberg


More information about the pve-devel mailing list