Solved netfront(netback) from 11.0 into 10.2

Hi!

We are running a number of ~1000 productive FreeBSD 10.2 DomUs on XenServer 6.5 and found that normal (ipv4, protocol mixture) network traffic is nearly 10(!) times slower compared to a similar FreeBSD 11.0 DomU on the same Dom0 (all outgoing traffic - not between the VMs).

Does anyone has an idea if the netfront devices from 11.0 could be included into a current 10.2 kernel?

We are unable to make give all productive VMs a complete world upgrade to 11.0 but we really suffer from the slow speed of in the FreeBSD 10.2 DomUs.

Is it netfront at all? The Dom0 here is the centos of cirtrix here.

Any help is really welcome and appreciated!!

Jimmy
 
Is this with vtnet(4) drivers? Have you tried turning TSO and/or LRO off? If I recall correctly there were some issues with it which causes packets to have an incorrect checksum and needed to be re-sent (which causes the slowness). I'm not entirely sure but I think I read somewhere this was fixed in later editions.

Besides jumping the gun and going for the, soon to be released, 11.0 have you tried 10.3?
 
SirDice Many thanks for your input! vtnet(4)() drivers are compiled into the kerne and the devices are xnX - so I guess this is correct. Is it enough to disable TSO4/LRO on the DomU? I will give it a try.
I will setup an 10.3 today as DomU to compare.

Is it possible at all to e.g. 'just' change the /boot/kernel like one can (in some cases) in linux? Or will this fail in any case? Or just taking over the netfront code? Just fishing around ;)
 
Is it enough to disable TSO4/LRO on the DomU?
Yes, that's where the issue is. You can try it "on-the-fly" using ifconfig vtnet0 -tso for example.

Is it possible at all to e.g. 'just' change the /boot/kernel like one can (in some cases) in linux? Or will this fail in any case?
On FreeBSD the kernel and userland (commonly known as "world") are a complete entity. Unlike Linux where the kernel is its own entity and can be swapped out for other versions.
 
Back
Top