12:42 a.m.
On 2013-06-12, at 7:58, "Prasad Surampudi"
<prasad.surampudi@theatsgroup.com<mailto:prasad.surampudi@theatsgroup.com>>
wrote:
We are doing a POC with Lustre Version 2.3. We have three VMs functioning as MDS servers
and the MGS is collocated with one of the MDS Server. MDS and OSS failovers using Redhat
Cluster Manger worked fine for us. I’m interested to know what happens to the cluster if
the MGS servers crashes. I have gone thru the Lustre manual
Presumably if you are on this list you are referring to the manual maintained by
Whamcloud/Intel?
and could not find any information to make the MGS server highly available like MDS and
OSS. So, given below are my questions:
1. From the Lustre manual it appears that we can have only one MGS server in a
cluster. Can someone confirm ? Is it possible to have multiple MGS servers with failover
using a Cluster Manager?
For maximum availability the MGS should also be configured with HA, and can be failed over
using Cluster Manager in the same way as the MDS. Systems are often configured with the
MGS on the backup MDS node so that separate hardware is not required.
If the MGS is configured in HA mode, then the clients should list both MGS nodes at mount
time, so that they will try both nodes to get the configuration.
2. What impact the Lustre clients will have if the MGS server goes down since we
reference MGS Server when the client mounts the Lustre file systems? Can the clients
continue the I/O since the MDS and ODS servers are still available?
If the MGS fails, it does not directly affect normal client operation. However, it would
prevent new clients or servers from mounting the filesystem, and any global configuration
changes (lctl conf_param ).
In Lustre 2.3, the MGS is also used for Imperative Recovery, which speeds up the recovery
process after a server crash. If the MGS is colocated with another server (e.g. the MDS)
and that node fails, then recovery will still complete, but will wait the full recovery
timeout as it did with older versions if Lustre.
3. Does the MGS server holds the completed Cluster configuration data? What is the
process for backing up MGS file system?
It holds the full Lustre configuration data. The easiest way to back it up is probably
"dd" of the whole device, since it is typically very small. The MGS content does
not change very often, only when configuration changes are made (add OSTs, conf_param) so
no need for daily backups (unless this is easier to not have to think about doing it
manually).
If the MGS is completely lost, the configuration can also be recreated on a new MGS using
"lctl --writeconf" on the servers, see the manual for details.
Cheers, Andreas