Closed GoogleCodeExporter closed 9 years ago
To see if it would help (solve) the problem we upgraded to version 2.9.3 of
ganeti, however it did not resolve this.
Greetings,
Eadric Wildeboer
Original comment by eadric.w...@gmail.com
on 4 Feb 2014 at 8:55
I just want to make sure, that this is not a duplicate of 694. How did you set
the spindle oversubscription ratio and how many instances do you have per node?
Original comment by aeh...@google.com
on 5 Feb 2014 at 8:55
Hi,
I hope the following info helps.
node01: 21 instances
node02: 20 instances
node03: 21 instances
node04: 19 instances
All instances are drbd with secondaries spread through the cluster in a
reasonable ballance
gnt-cluster info | grep spindle
spindle_count: 1
spindle-use: 12
spindle-use: 1
spindle-use: 1
spindle-ratio: 32
Original comment by eadric.w...@gmail.com
on 5 Feb 2014 at 11:23
So, unless I'm misunderstanding something,
each node has one spindle, allowing a 32 to 1 oversubscription.
about 20 instances per node, with drbd should give about 40 spindles
used per node... That would at least explain the problem.
Does 'gnt-cluster modify --ipolicy-spindle-ratio=128' solve the problem?
Original comment by aeh...@google.com
on 5 Feb 2014 at 11:48
Ok,
I've done that and it does help. Now I get messages about CPU. I'm gathering
this is to be seen as expected behavior.
I have two questions:
Is there documentation about the best practice for tuning these settings to the
available hardware? I believe these are default settings.
Is this behavior new sinds 2.5? This same cluster run under 2.5 in the past
with about twice the instances as it has now and I don't recollect that we
experienced these issues.
Original comment by eadric.w...@gmail.com
on 5 Feb 2014 at 3:54
Hi,
We currently have no documentation on the proper values unfortunately. The
reason we made it possible to toggle them is indeed that we didn't know of a
"correct" value that would work for everyone. The right value depends more on
the workload on the instances, and how much they interfere with each other than
on the actual hardware.
Policies IIRC have been added in 2.6+, so that one can avoid inadvertently
oversubscribe a node.
Thanks,
Guido
Original comment by ultrot...@google.com
on 5 Feb 2014 at 4:02
hi,
Fair enough.
Perhaps there could be a forum thread or something with experiences from users
so that over time some trends could surface.
Should I start one?
Original comment by eadric.w...@gmail.com
on 5 Feb 2014 at 4:30
I think this is actually a duplicate of Issue 503, although the headline of
that one is missleading (will fix this).
Original comment by hel...@google.com
on 12 Feb 2014 at 3:50
Original issue reported on code.google.com by
eadric.w...@gmail.com
on 4 Feb 2014 at 7:36