[sheepdog-users] Guests get stuck when doing a lot of small io

[Thomas Fjellstrom]

I've setup a mostly working sheepdog single backend "cluster" on a NAS, with a 
secondary gateway node on a vm host.

Performance seems reasonable in general, but running a iozone -a -s 8G on a 
guest allocated 4G ram will lock up iozone and a jbd2 kernel thread for 2 or 
more minutes when testing 4k block sizes.

It tends not to happen with larger block sizes.

My network is plain GbE copper with a semi dedicated network for the data 
traffic (almost nothing else is using that lan segment. its on its own 24 port 
netgear, and the hosts have dedicated intel nics). Only "Services" related 
traffic uses that lan, and it is connected to the internet and my main lan 
through a pfsense appliance. Network performance is pretty darn good. Can 
easily get ~100MBps across each lan segment. NAS disk performance is 300MBps+ 
on a raid 5, vm host box is configured to use a local cache which it doesn't 
seem to hit (I presume because the vm volume isn't set to writeback). 

Sheepdog data throughput from host to nas gets upwards of 80MBps+ which I am 
not upset with at all. Seems pretty decent.

Thanks.

-- 
Thomas Fjellstrom
thomas at fjellstrom.ca