[LU-13641] socklnd: remove use_tcp_bonding option in favor of LNet Multi-Rail

Related — Sat, 17 Jun 2023 05:38:46 +0000

TCP bonding in socklnd over-complicates the code and there is no evidence it's being used anywhere. With LNet Multi-Rail, the use_tcp_bonding option has become obsolete. Add a deprecation message for earlier releases. Remove it in the 2.15 release.

Multi-Rail feature doesn't need to be explicitly enabled. To use MR instead of the use_tcp_bonding configuration option, group the interfaces on the same network using the lnetctl utility:

lnetctl net add --net tcp --if eth0,eth1

or via the /etc/modprobe.d/lnet.conf or /etc/modprobe.d/lustre.conf configuration file:

options lnet networks="tcp(eth0,eth1)"

and make sure dynamic discovery is enabled:

lnetctl set discovery 1

MR will aggregate the throughput of all available networks/interfaces shared between peer nodes.

See LNet Software Multi-Rail Configuration in the Lustre Operations Manual for more details.

Whamcloud Community JIRA

[LU-13641] socklnd: remove use_tcp_bonding option in favor of LNet Multi-Rail