Lustre discuss - Mar 2010 - networking problem between OSS and MGS

Hi,

I am trying to setup the network between my Lustre servers, actually I am not 
sure whether this will work at all what I am trying, so here it goes.
I use SLES 11 x86_64 and Lustre 1.8.2 on all hosts.
As a start, I have a separate MGS (192.168.0.150) in one network, and a 
separate MDS (192.168.1.216) in a second network, and two OSS nodes in a third 
network. The reason to have the stuff in different network is to have a 
firewall in between, that can restrict access to the hosts. 

The MGS and MDS hosts are each connected with 1 GB Ethernet interface, the OSS 
nodes are connected with two GB network interfaces to the switch in the same 
network. OSS node 1: 192.168.2.21 and 192.168.2.121 OSS node 2: 192.168.2.22 
192.168.2.122.
Filesystems were created on the MDS, MDT.
The filesystems on the OSS nodes were first created the following way:
On OSS1:
mkfs.lustre --fsname=WB01 --ost --mgsnode=192.168.0.150 at tcp 
/dev/mapper/WBOSS1_part1
mkfs.lustre --fsname=WB01 --ost --mgsnode=192.168.0.150 at tcp 
/dev/mapper/WBOSS1_part2
On OSS2:
mkfs.lustre --fsname=WB01 --ost --mgsnode=192.168.0.150 at tcp 
/dev/mapper/WBOSS2_part1
mkfs.lustre --fsname=WB01 --ost --mgsnode=192.168.0.150 at tcp 
/dev/mapper/WBOSS2_part2

and on the OSS hosts, I tried to put the following into the 
/etc/modprobe.conf.local:
options lnet networks=tcp0(eth0,eth1)

However, then only the eth0 was used, the interface where the default route is 
configured.

Then I tried to do the following:
On OSS1:
mkfs.lustre --fsname=WB01 --ost --mgsnode=192.168.0.150 at tcp0 
/dev/mapper/WBOSS1_part1
mkfs.lustre --fsname=WB01 --ost --mgsnode=192.168.0.150 at tcp1 
/dev/mapper/WBOSS1_part2
On OSS2:
mkfs.lustre --fsname=WB01 --ost --mgsnode=192.168.0.150 at tcp0 
/dev/mapper/WBOSS2_part1
mkfs.lustre --fsname=WB01 --ost --mgsnode=192.168.0.150 at tcp1 
/dev/mapper/WBOSS2_part2

and configured this in the /etc/modprobe.conf.local:
options lnet networks=tcp0(eth1),tcp1(eth0)
Then /dev/mapper/WBOSS1_part1 was mountable, but when I tried to mount 
/dev/mapper/WBOSS1_part2 then I get this error message:
mount.lustre: mount /dev/mapper/WBOSS1_part2 at /lustre/WBOSS2-1 failed: 
Cannot send after transport endpoint shutdown
and on the MGS in dmesg I see:
LustreError: 120-3: Refusing connection from 192.168.2.21 for 
192.168.0.150 at tcp1: No matching NI
Where the 192.168.2.21 is the IP of eth0 (tcp1)
when I switch the default routes, then the problem persists

I also tried to configure /etc/modprobe.conf.local the other way around:
options lnet networks=tcp0(eth0),tcp1(eth1)
Then I also was able to mount /dev/mapper/WBOSS1_part1, but not 
/dev/mapper/WBOSS1_part2, with the same error, It just takes the other
Ethernet interface.

Also lctl ping only works for one interface:
oss1:~ # lctl ping 192.168.0.150 at tcp0
12345-0 at lo
12345-192.168.0.150 at tcp
oss1:~ # lctl ping 192.168.0.150 at tcp1
failed to ping 192.168.0.150 at tcp1: Input/output error

Looking further into the documentation, I saw that some kind of load balancing 
can be setup with ip2nets parameter in /etc/modprobe.conf.local
But as far as I can see, I cannot specify two interface to be used to 
communicate with the MGS?

So is above setup possible at all? I hope I was clear enough with my 
explanation ;)

When I now try to to setup bonding on the OSS hosts, which bonding mode would 
be the preferred one to choose?
The manual only talks about the recommended xmit_hash_policy, but does not 
recommend a bonding mode. The switch is capable of doing 802.1ad trunking.

And a more general question: What is recommended anyways, bonding the two 
available interfaces on the OSS servers, or getting Lustre and the routing 
configured to use two separate network interfaces?

regards,
Sebastian