[sheepdog-users] Issues with gateway mode

Gerald Richter richter at ecos.de
Mon Nov 9 06:43:32 CET 2015

Hi Hitoshi,

as long as the object cache doesn't work relieable my setup doesn't make 
sense, because of the slow network connections (only 200MBit/s).

This will only work if most reads and writes are statisfied by the local 
object cache. If every read and write goes through the network, than 
performance will be very poor.

At the moment I have only 2 storage nodes and 3 gateways with 6 VMs for 
testing (and object cache turned on). This works fine, but as you say 
that object cache should not be used, this setup will not scale (in the 
end there will be about 100 VMs on 8-10 nodes).

I need something that allows me such a setup over slow network 
connections. This mean that must be some cacheing and an asynchronous 
write back. For the kind of VMs I am running it's ok, if some data is 
lost in case of failure.

Sheepdog with object cache looked very promising for such a setup. I 
didn't found any other solution so far.



Am 09.11.2015 um 05:01 schrieb Hitoshi Mitake:
> Hi Gerald,
> On Sat, Nov 7, 2015 at 9:05 PM, Gerald Richter <richter at ecos.de> wrote:
>> Hi,
>> I need to setup a storage cluster over slow network connections.
>> Sheepdog in gateway mode seems to be the ideal solution for me. That mean I
>> have three or more storage nodes and a lot of gateway node that only uses
>> the sheepdog cache. As long as most request are statisfied by the cache
>> (which is an ssd), performance should be ok. Slow network connection aren't
>> a problem as long as the cache uses write back mode.
>> I have setup a test environment with 0.9.2 . As long as the vm are small
>> enough to fit completly in the cache it seems to work, but when the vm is
>> bigger than the cache I get I/O errors after some time. Also after a reboot
>> of the host machine it seems to take very long time, until I can boot the
>> vm, also the cache is still uptodate.
>> Browseing through the mailing list archive I found
>> http://lists.wpkg.org/pipermail/sheepdog-users/2015-August/015120.html,
>> where Hitoshi said "object cache is not maintained currently. Please do not
>> use it.". Is this still true? or is there any way to use the setup I
>> described above?
> Yes, it's true. Please don't use object cache.
> BTW, many gateways + fewer storage nodes is dangerous. Because the
> gateways doesn't contribute to increasing parallelism of disk I/O. If
> a number of VMs and storage nodes are unbalanced, performance of
> virtual disks will be very bad. Could you provide a detailed number of
> VMs, gateway nodes, and storage nodes?
> Thanks,
> Hitoshi
>> Thanks & Regards
>> Gerald
>> --
>> sheepdog-users mailing lists
>> sheepdog-users at lists.wpkg.org
>> https://lists.wpkg.org/mailman/listinfo/sheepdog-users
> !DSPAM:416,56401a9a23791711617062!

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.wpkg.org/pipermail/sheepdog-users/attachments/20151109/26435e99/attachment.html>

More information about the sheepdog-users mailing list