The strict ordering of MGT, MDT, OST is only a hard requirement for the very first start
of the file system, because the targets need to compete the initial registration with the
MGT before they can be used. Subsequent cold restarts of the filesystem don’t necessarily
need to start any particular order, but it’s still not a bad idea to start MGT first since
all the servers will attempt to connect to it when they start, though they should start
without it eventually.
On May 23, 2014, at 09:16 , White, Cliff <cliff.white(a)intel.com> wrote:
In a failover situation, any target can be stopped and restarted
impact on other nodes. The startup order in the manual is for a cold
startup/full shutdown situation, and does not apply to a running
filesystem and failover.
You should not have the ordering directive, I think. In particular, the
MGT is only used for client/server mounts, if the filesystem is up and
running the MGT failover should be very transparent to the rest of the
cluster and should never require another node to be restarted.
On 5/23/14, 7:58 AM, "Riccardo Murri" <riccardo.murri(a)uzh.ch> wrote:
> The online Lustre manual recommends that Lustre targets are started in
> this order: MGT, MDT, OSTs, clients.
> Now we are setting up an HA cluster with Pacemaker, and a strict
> ordering directive ("order ... Mandatory: ostXX mdt mgt") results in a
> complete restart of all targets if the MGT is migrated. In my
> experience (with Lustre 1.8.5) this is most of the time unnecessary,
> and Lustre can recover from a single target restart. However, we have
> recently switched to Lustre 2.4.3 and things might have changed.
> So the question is: is this order strict (in Lustre 2.4.3), or can a
> target be stopped and restarted on another node without affecting the
> targets running on other nodes?
> Thanks for any help!
> Riccardo Murri
> Grid Computing Competence Centre
> University of Zurich
> Winterthurerstrasse 190, CH-8057 Zürich (Switzerland)
> Tel: +41 44 635 4222
> Fax: +41 44 635 6888
> Lustre-discuss mailing list
HPDD-discuss mailing list