Re: [Xen-users] High availability Xen with bonding

[Dirk <dirk.schulz@xxxxxxxxxxxxx>]

 Eric,

Am 18.10.10 08:11, schrieb Eric van Blokland:
> Hey Florian,
>
> In my reply to Bart I've explained I'm having some connectivity issues. I'm not 
> using VLANs either, nor have access to expensive switches with link aggregation 
> support. I'm going to have a look at the manuals you mentioned. Perhaps there 
> is some setting I missed which is causing my issues. I'll keep you all updated 
> on my progress.
I am using bonding with xen bridges on CentOS and Debian without LACP or 
VLANs. If you could describe your connection issues I can help you, maybe.
Dirk
> Regards,
>
> Eric
>
> -----Oorspronkelijk bericht-----
> Van: Florian Heigl [mailto:florian.heigl@xxxxxxxxx]
> Verzonden: zondag 17 oktober 2010 7:19
> Aan: Bart Coninckx
> CC: xen-users@xxxxxxxxxxxxxxxxxxx; Eric van Blokland
> Onderwerp: Re: [Xen-users] High availability Xen with bonding
>
> Hi both,
>
> I had very good success after some pulling-hairs.
> I run lacp + vlan trunking.
> Key assumptions:
> - All device setup (eth, bond, bridges) is done via normal OS config,
> because that is more reliable.
> - All libvirt stuff is disabled, it just limits Xen's possibilities to
> "home user level" by assuming you'd only have one bridge.
> (chkconfig  XXX off ...)
> - No messing with ARP is wanted
> - You have switches current enough to do "real" LACP
>
>
> There's a very good (and I think the only working one) manual in the
> Oracle VM wiki at
> http://wiki.oracle.com/page/Oracle+VM+Server+Configuration-+bonded+and+trunked+network+interfaces
>
> I myself had followed one manual from redhat, which left off somewhere
> in the middle.
> It's called "Xen_Networking.pdf" by Mark Nielsen. It's a good intro,
> but only covers 50% of a good setup.
>
> Notes:
> a) if you look not just at link aggregation but a VLAN-heavy
> environment there might be a point (>128 VLANs) where the number of
> virtual bridges might become an issue. Then wait for OpenVswitch to
> mature or email xen-devel and ask for the status of "vnetd". (just
> kidding)
> b) using ethernet (n ethernet links into bond0) and infiniband (2
> infiniband hca ports into bond1) bonding on the same host is more
> tricky. it seems the ethernet bonding driver tries to cover infiniband
> too. The setup is completely undocumented. It is possible, but when I
> tried it just didn't pass any more traffic.
> For the setup check the following thread in HP itrc:
> http://forums13.itrc.hp.com/service/forums/questionanswer.do?admit=109447627+1287292335750+28353475&threadId=1445752
> c) added speed is only guaranteed for multiple connections. if you do
> it via bonding, you need a lacp algorithm in your switch that will
> hash based on the ip destination ports, not just mac address or ip
> address. current cisco gear can do that. for plain iscsi your path
> grouping would decide if you see loadbalancing with multiple iSCSI
> lans.
>
>
> Hope you get it to work!
>
>
>
> Florian
>
> 2010/10/16 Bart Coninckx<bart.coninckx@xxxxxxxxxx>:
> > On Friday 15 October 2010 13:44:42 Eric van Blokland wrote:
> > > Hello everyone,
> > >
> > > A few days back I decided to give Ethernet port bonding in Xen another try.
> > > I've never been able to get it to work properly and after a short search I
> > > found the network-bridge-bonding script shipped with CentOS-5 probably
> > > wasn't going to solve my problems. Instead of searching for a tailored
> [...]
> > curious to see your progress in this. Up till now I tackled network redundancy
> > with multipathing, not with bonding. However, this does not provide added
> > speed, though theoretically it should. So I recently decided to switch to
> > bonding for the hypervisors in their connections to iSCSI, using rr and
> > running over seperate switches, just like you but I'm not at the point of
> > installing domU's, so I can't really comment on how and if it works. Will know
> > next week though so I will return to this post with my findings...
> that should definitely get you increased speed, plus multiple iSCSI
> connections via separate subnets / nics is the only way you can get
> close to FC reliabilty for lower budget.
_______________________________________________
Xen-users mailing list
Xen-users@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-users