[sheepdog-users] Concern about sheepdog performance

[Liu Yuan]

On 12/18/2012 11:43 AM, MORITA Kazutaka wrote:
> At Tue, 18 Dec 2012 11:21:31 +0800,
> Liu Yuan wrote:
>>
>> On 12/18/2012 11:11 AM, Jan Dvořák wrote:
>>> Recovery of lost nodes we tested with a similar setup was painfully
>>> slow, however.
>>>
>>> Take this with a grain of salt, it's been too long.
>>
>> Recovery is slow because we don't want recovery process saturate the
>> network. Serving in-flight IO during recovery is preferred over
>> recovery. I think with this strategy, VMs will feel more smooth disk IO
>> during recovery process. Also, there is always conflict between data
>> balance and minimal data migration. We prefer data balance, so this
>> preference will result in more data migration than necessary.
> 
> I think it would be a good idea to add support for multiple NICs and
> use a dedicated NIC for object recovery.  Then, we don't have to worry
> about network saturation caused by recovery.
> 

Yes, I think membership membership heartbeat, notify event and recovery
data transfer can go one NIC and VM disk IO go on another. This would
improve the scalability, reliability and provide possible infrastructure
to further speedup recovery.

Thanks,
Yuan