[ovirt-users] Multipath handling in oVirt

[Yura Poltoratskiy]

Hi,

As for me personally I have such a config: compute nodes with 4x1G nics 
and storages with 2x1G nics and 2 switches (not stackable). All servers 
runs on CentOS 7.X (7.3 at this monent).

On compute nodes I have bonding with two nic1 and nic2 (attached to 
different switches) for mgmt and VM's network, and the other two nics 
nic3 and nic4 without bonding (and also attached to different switches). 
On storage nodes I have no bonding and nics nic1 and nic2 connected to 
different switches.

I have two networks for iSCSI: 10.0.2.0/24 and 10.0.3.0/24, nic1 of 
storage and nic3 of computes connected to one network; nic2 of storage 
and nic4 of computes - to another one.

On webUI I've created network iSCSI1 and iSCSI2 for nic3 and nic4, also 
created multipath. To have active/active links with double bw throughput 
I've added 'path_grouping_policy "multibus"' in defaults section of 
/etc/multipath.conf.

After all of that, I have 200+MB/sec throughput to the storage (like 
raid0 with 2 sata hdd) and I can lose one nic/link/swith without 
stopping vms.

[root at compute02 ~]# multipath -ll
360014052f28c9a600000000000000000 dm-6 LIO-ORG ,ClusterLunHDD
size=902G features='0' hwhandler='0' wp=rw
`-+- policy='service-time 0' prio=1 status=active
   |- 6:0:0:0 sdc 8:32  active ready running
   `- 8:0:0:0 sdf 8:80  active ready running
36001405551a9610d09b4ff9aa836b906 dm-40 LIO-ORG ,SSD_DOMAIN
size=915G features='0' hwhandler='0' wp=rw
`-+- policy='service-time 0' prio=1 status=active
   |- 7:0:0:0 sde 8:64  active ready running
   `- 9:0:0:0 sdh 8:112 active ready running
360014055eb8d30a91044649bda9ee620 dm-5 LIO-ORG ,ClusterLunSSD
size=135G features='0' hwhandler='0' wp=rw
`-+- policy='service-time 0' prio=1 status=active
   |- 6:0:0:1 sdd 8:48  active ready running
   `- 8:0:0:1 sdg 8:96  active ready running

[root at compute02 ~]# iscsiadm -m session
tcp: [1] 10.0.3.200:3260,1 iqn.2015-09.lab.lnx-san:storage (non-flash)
tcp: [2] 10.0.3.203:3260,1 iqn.2016-10.local.ntu:storage3 (non-flash)
tcp: [3] 10.0.3.200:3260,1 iqn.2015-09.lab.lnx-san:storage (non-flash)
tcp: [4] 10.0.3.203:3260,1 iqn.2016-10.local.ntu:storage3 (non-flash)

[root at compute02 ~]# ip route show | head -4
default via 10.0.1.1 dev ovirtmgmt
10.0.1.0/24 dev ovirtmgmt  proto kernel  scope link  src 10.0.1.102
10.0.2.0/24 dev enp5s0.2  proto kernel  scope link  src 10.0.2.102
10.0.3.0/24 dev enp2s0.3  proto kernel  scope link  src 10.0.3.102

[root at compute02 ~]# brctl show ovirtmgmt
bridge name     bridge id               STP enabled     interfaces
ovirtmgmt               8000.000475b4f262       no bond0.1001

[root at compute02 ~]# cat /proc/net/bonding/bond0 | grep "Bonding\|Slave 
Interface"
Ethernet Channel Bonding Driver: v3.7.1 (April 27, 2011)
Bonding Mode: fault-tolerance (active-backup)
Slave Interface: enp4s6
Slave Interface: enp6s0


01.02.2017 12:50, Nicolas Ecarnot пишет:
> Hello,
>
> I'm starting over on this subject because I wanted to clarify what was 
> the oVirt way to manage multipathing.
>
> (Here I will talk only about the data/iSCSI/SAN/LUN/you name it 
> networks.)
> According to what I see in the host network setup, one can assign 
> *ONE* data network to an interface or to a group of interfaces.
>
> That implies that if my host has two data-dedicated interfaces, I can
> - either group them using bonding (and oVirt is handy for that in the 
> host network setup), then assign the data virtual network to this bond
> - either assign each nic a different ip in each a different VLAN, then 
> use two different data networks, and assign them each a different data 
> network. I never played this game and don't know where it's going to.
>
> At first, may the oVirt storage experts comment on the above to check 
> it's ok.
>
> Then, as many users here, our hardware is this :
> - Hosts : Dell poweredge, mostly blades (M610,620,630), or rack servers
> - SANs : Equallogic PS4xxx and PS6xxx
>
> Equallogic's recommendation is that bonding is evil in iSCSI access.
> To them, multipath is the only true way.
> After reading tons of docs and using Dell support, everything is 
> telling me to use at least two different NICs with different ip, not 
> bonded - using the same network is bad but ok.
>
> How can oVirt handle that ?
>