How do I configure BeeGFS to do native IB rather than IP over IB?
Preliminary: BeeGFS Installation
By default BeeGFS is not installed on Bright Cluster Manager. Setting it up is straightforward; cm-beegfs-setup works as described in the administrator manual:
Configuration Of BeeGFS Native IB Support
The following steps rely on the BeeGFS documentation at https://www.beegfs.io/wiki/NativeInfinibandSupport25
After cm-beegfs-setup installation is finished, communication between management BeeGFS elements defaults to internalnet, as can be verified by running the commands:
# beegfs-ctl --listnodes --nodetype=storage --details
# beegfs-ctl --listnodes --nodetype=meta --details
# beegfs-ctl --listnodes --nodetype=client --details
For BeeGFS version 7.1 and above, BeeGFS communications can be made to switch over to the IB interface as follows:
should be edited.
should be changed to:
The package libbeegfs-ib should be installed into the image that is used by the BeeGFS nodes:
# chroot /cm/images/default-image
# yum install libbeegfs-ib
At this point, the beegfs-ctl commands that were run earlier on in this article should output that BeeGFS is using the IB interface:
[root@goofy default-image]# beegfs-ctl --listnodes --nodetype=meta --details
node1 [ID: 1]
Ports: UDP: 8005; TCP: 8005
Interfaces: ib0(RDMA) br0:vxlan(TCP) br0(TCP) ib0(TCP)
The text "RDMA" here means that the associated interface is enabled for the native Infiniband protocol (IB verbs).
Additional configuration: disabling the ibacm service
A typical source of trouble having the ibacm service (/etc/init.d/ibacm) still running on the machines. This service causes RDMA connection attempts to stall. It should be disabled in all nodes:
# systemctl stop ibacm.service
# systemctl disable ibacm.service
- More configuration examples can be seen at: https://www.beegfs.io/wiki/NativeInfinibandSupport#hn_59ca4f8bbb_421
- In an RDMA-capable cluster, there may still be some BeeGFS communication (especially communication with the management service, which is not performance-critical) that still uses TCP/IP and UDP/IP transfer. On some hardware the default "connected" IP-over-IB mode of InfiniBand and Omni-Path does not seem to work well and results in spurious problems. If that seems to be the case, then switching the IPoIB mode to "datagram" on all hosts should be tried.