Setting MTU on Exalogic
- by csoto
For many reasons, a system administrator may want to change the MTU settings of a server. But in a system like Exalogic which contains lots of interconnected nodes and other various components, it's important to understand how this applies to the different networks.
For example, when bringing up bonding of InfiniBand an error like the following may be thrown:
Bringing up interface bond1: SIOCSIFMTU: Invalid argument
Both scripts ifcfg-ib0 and ifcfg-ib1 (from the /etc/sysconfig/network-scripts/ direectory) have MTU set to 65500, which is a valid MTU value only if all IPoIB slaves operate in connected mode and are configured with the same value, so the line below must be added to both network scripts and then restart the network:
CONNECTED_MODE=yes
By the way, an error of the form “SIOCSIFMTU: Invalid argument” indicates that the requested MTU was rejected by the kernel. Typically this would be due to it exceeding the maximum value supported by the interface hardware. In that case you must either reduce the MTU to a value that is supported or obtain more capable hardware. This problem has been seen when trying to modify the MTU using the ifconfig command, like the output of the example below:
[root@elxxcnxx ~]# ifconfig ib1 mtu 65520
SIOCSIFMTU: Invalid argument
It's important to insist that in most cases the nodes must be rebooted after the MTU size has been changed. Although in some circumstances it may work without a reboot, it is not how it is typically documented.
Now, in order to achieve a reduced memory consumption and improve performance for network traffic received on IPoIB related interfaces, it is recommend to reduce the MTU value in interface configuration files for IPoIB related bonds from 65520 to 64000. The change needs to be made to interface configuration files under the /etc/sysconfig/network-scripts directory and applies to the interface configuration files for bonds over IPoIB related slave devices, for example /etc/sysconfig/network-scripts/ifcfg-bond1. However, keep in mind that the numeric portion of the interface filenames that corresponding to IPoIB interfaces is expected to vary across compute nodes and vServers and so cannot be relied upon to identify which interface files are for bonds are over IPoIB rather than EoIB related slave interfaces.
To fix these MTU values to the recommended settings, there are very useful instructions and a script on the MOS Note 1624434.1, and it's applicable physical and virtual configurations of Exalogic.
Regarding the recommended MTU value for EoIB related interfaces, its maximum appropriate value is 1500.
If for some reason a vServer has been created with a higher value (set on the /etc/sysconfig/network-scripts/ifcfg-bond0 file), then it must be fixed. An error like the following could be thrown under this circumstance:
[root@vServer ~]# service network restart
...
Bringing up interface bond0: SIOCSIFMTU: Invalid argument
Also an error like the one below can be seen on the /var/log/messages file of the vServer:
kernel: T5074835532 [mlx4_vnic] eth1:vnic_change_mtu:360: failed: new_mtu 64000 2026
The MOS Note 1611657.1 is very useful for this purpose.
