[Sheepdog] Performance of sheepdog
Piavlo
piavka at cs.bgu.ac.il
Thu Jan 7 10:58:22 CET 2010
Piavlo wrote:
> Hi,
>
> PCextreme B.V. - Wido den Hollander wrote:
>
>> Hi,
>>
>> I have been following sheepdog's mailinglist ever since i saw a newspost
>> on www.linux-kvm.com, but what kept me wondering was the performance.
>>
>>
> I have not done any performance testing yet, mainly since the sheepdog
> patched kvm does not work for me.
> So currently I can only test image creation performance.
>
> But if disk performance is important for you probably need to separate
> the host running the KVM VMs completely
> from the VMs image storage device (like done with iSCSI or AOE but not
> DRBD) since otherwise the KVM host will probably be throttled
> by random disk access generated by your VMs (unless you have just a few
> VMs) plus occasional overhead while new images are being created.
>
> AFAIU with current sheepdog implementation the KVM host MUST be part of
> the sheepdog storage cluster.
> So it would be nice if sheepdog client and sheepdog server could be run
> on different hosts - currently one can overcome this
> limitation by keeping the btrfs system not on local disk but on
> iSCSI/AOE or any other SAN device.
> But addition of another SAN technology complicates the overall solution
> and reduces stability.
>
On a second thought using using KVM host as iSCSI initiator for
underlying btrfs device, while
the iSCSI target is exported from properly tuned ZFS for the workload
on Opensolaris on a host with a lots of RAM
would probably give you much better performance than a separated
sheepdog server on the same host instead of Opensolaris with ZFS.
> So if sheepdog could do the job alone it would be much better.
>
>> At the moment i don't have 3 servers with SMX/VMX to spare for testing
>> sheepdog, so i haven't been able to test sheepdog jet.
>>
>> At the company i work we use a lot of KVM, our redundant storage is
>> build on top of iSCSI and DRBD. The performance is great (IOps), so
>> that's not an issue, but the waste of hardware is (Master/Slave setup).
>>
>>
>>
> AFAIU you have combined iSCSI with DRBD , thus the Master/Slave comes
> from DRBD.
> With sheepdog you still have hardware overhead since you need to run at
> least two sheepdog
> server hosts - which for performance reasons should be separated from
> the sheepdog clients (the KVM hosts).
>
> Alex
>
>> The question that kept me busy was, how is the performance of sheepdog?
>> Yes, i understand that sheepdog and btrfs are still under heavy
>> development, but in this stage, how is it performing?
>>
>> Since sheepdog really seems THE solution for building a real cloud with
>> KVM.
>>
>> I'll keep following sheepdog closely and try to set-up a test
>> envirioment asap!
>>
>>
>>
>
>
More information about the sheepdog
mailing list