[OmniOS-discuss] horrible disk performance under KVM

[Narayan Desai]

Picking up on an earlier point, it looks like both virtio net and virtio
block device support have been integrated into mainline illumos for a
while, so in theory it should work, though it might take some doing to get
a previously installed system to switch to virtio.

Has anyone actually tried this?
 -nld


On Fri, Sep 13, 2013 at 6:38 PM, ashley willis <ashleywillis76 at gmail.com>wrote:

> I decided to run the mongorestore tool from the bare metal OmniOS to
> the mongo server running in KVM CentOS, and it is running much quicker
> -- 50% in 12m52s (and the CPU and memory for mongorestore is only
> 0.3%). MongoDB is not using much CPU in the VM, either. Total import
> time this way is 32m12.235s -- still much slower than just bare metal,
> but not insanely slow like before (the NFS->NFS all within the OmniOS
> VM was 92m48s). Giving the VM 6 cores and 14GB of ram (instead of 2
> and 2) gave somewhat better performance: 27m18.305s. Going back to
> if=ide, 35m51.575s, so if=virtio does help here some.
>
> Is there anything else that can be done to improve disk I/O in a VM,
> short of giving the VM an entire physical disk?
>
> -ashley
>
> On Fri, Sep 13, 2013 at 4:58 PM, ashley willis <ashleywillis76 at gmail.com>
> wrote:
> > So, a CentOS 6.4 VM is still horrible with virtio -- it's loaded 2% of
> > the database in 6 minutes onto a 30g ext4 partition.
> >
> > On Fri, Sep 13, 2013 at 4:24 PM, ashley willis <ashleywillis76 at gmail.com>
> wrote:
> >> that works for booting centos 6.4, but not for booting omnios -- grub
> >> comes up but it keeps looping back to the grub menu. granted, i don't
> >> really need to boot omnios in kvm. i had horrible performance in
> >> centos as well, so i will try this.
> >>
> >> for comparison to my home machine, on a work server i loaded that same
> >> dataset in 4m10s (xeon 2*6 cores, 60GB ram, 12 600GB 10k SAS drives in
> >> raid10) reading from disks in a different box.
> >>
> >> -ashley
> >>
> >> On Fri, Sep 13, 2013 at 3:50 PM, Narayan Desai <narayan.desai at gmail.com>
> wrote:
> >>> Depending on the availability of the guest-side drivers, it should be
> a lot
> >>> faster to use virtio, both for disk and net.
> >>>
> >>> Does anyone know if virtio support ships now in OmniOS? I know there
> have
> >>> been some drivers floating around for a while now, but haven't tried
> them.
> >>>  -nld
> >>>
> >>>
> >>> On Fri, Sep 13, 2013 at 1:45 PM, ashley willis <
> ashleywillis76 at gmail.com>
> >>> wrote:
> >>>>
> >>>> I have 14GB dump of a MongoDB database that took 10 minutes to restore
> >>>> on bare metal OmniOS -- I'm fine with this (system is an 8-core AMD
> >>>> FX-8150, 32GB ram, pair of 500GB WD 'blue" drives mirrored, 90GB ssd
> >>>> cache -- all desktop-grade).
> >>>>
> >>>> Running OmniOS under KVM on this box, the same operation took 246
> >>>> minutes with the dump available over an NFS share from the bare metal
> >>>> system (given 2 cores, 2GB ram, and a 30GB zvol).
> >>>>
> >>>> Under the same KVM instance, I decided to see the performance reading
> >>>> the dump from NFS and having the MongoDB data directory on this same
> >>>> NFS share, and it's on track to finish in under 60 minutes (half-way
> >>>> done). That's more than 4 times faster.
> >>>>
> >>>> What gives?
> >>>>
> >>>> /usr/bin/qemu-system-x86_64 -name OmniOS_Text_r151006p.iso -enable-kvm
> >>>> -vnc 0.0.0.0:15 -smp 2 -m 2048 -no-hpet -localtime -drive
> >>>> file=/dev/zvol/rdsk/rpool/omnios,if=ide,index=0 -net
> >>>> nic,vlan=0,name=net0,model=e1000,macaddr=2:8:20:e:15:33 -net
> >>>> vnic,vlan=0,name=net0,ifname=vnic10,macaddr=2:8:20:e:15:33 -vga std
> >>>> -daemonize
> >>>>
> >>>> -ashley
> >>>> _______________________________________________
> >>>> OmniOS-discuss mailing list
> >>>> OmniOS-discuss at lists.omniti.com
> >>>> http://lists.omniti.com/mailman/listinfo/omnios-discuss
> >>>
> >>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://omniosce.org/ml-archive/attachments/20130913/75899559/attachment.html>