Could you create a tracker for this and attach an
osdmap as well as
some recent balancer output (perhaps at a higher debug level if
possible)?
There are some improvements awaiting backport to nautilus for the
C++/python interface just FYI [0]
You might also look at gathering output using something like [1] to
try to narrow down further what is causing the high CPU consumption.
[0]
On Fri, May 8, 2020 at 1:10 AM Andras Pataki
<apataki(a)flatironinstitute.org> wrote:
Hi everyone,
After some investigation, it looks like on our large cluster, ceph-mgr
is not able to keep up with the status updates from about 3500 OSDs. By
default OSDs send updates to ceph-mgr every 5 seconds, which, in our
case, turns to about 700 messages/s to ceph-mgr. It looks from gdb
traces that ceph-mgr runs some python code for each of them - so 700
python snipets/s might be too much. Increasing mgr_stats_period to 15
seconds reduces the load and brings ceph-mgr back to responsive again.
Unfortunately this isn't sustainable since if we were to expand the
cluster, we'd need to further reduce the update frequency from OSDs.
I also checked our other clusters and they have about proportionately
lower load on ceph-mgr based on their OSD counts.
Any thoughts about the scalability of ceph-mgr to a large number of
OSDs? We recently upgraded this cluster from Mimic, where we didn't see
this issue.
Andras
On 5/1/20 8:48 AM, Andras Pataki wrote:
Also just a follow-up on the misbehavior of
ceph-mgr. It looks like
the upmap balancer is not acting reasonably either. It is trying to
create upmap entries every minute or so - and claims to be successful,
but they never show up in the OSD map. Setting the logging to
'debug', I see upmap entries created such as:
2020-05-01 08:43:07.909 7fffca074700 4 mgr[balancer] ceph osd
pg-upmap-items 9.60c4 mappings [{'to': 3313L, 'from': 3371L}]
2020-05-01 08:43:07.909 7fffca074700 4 mgr[balancer] ceph osd
pg-upmap-items 9.632b mappings [{'to': 2187L, 'from': 1477L}]
2020-05-01 08:43:07.909 7fffca074700 4 mgr[balancer] ceph osd
pg-upmap-items 9.6b9c mappings [{'to': 3315L, 'from': 3371L}]
2020-05-01 08:43:07.909 7fffca074700 4 mgr[balancer] ceph osd
pg-upmap-items 9.6bf6 mappings [{'to': 1581L, 'from': 1477L}]
2020-05-01 08:43:07.909 7fffca074700 4 mgr[balancer] ceph osd
pg-upmap-items 9.7da4 mappings [{'to': 2419L, 'from': 2537L}]
...
2020-05-01 08:43:07.909 7fffca074700 20 mgr[balancer] commands
[<mgr_module.CommandResult object at 0x7fffcc990550>,
<mgr_module.CommandResult object at 0x7fffcc990fd0>,
<mgr_module.CommandResult object at 0x7fffcc9907d0>, <mgr_module.Com
mandResult object at 0x7fffcc990650>, <mgr_module.CommandResult object
at 0x7fffcc990610>, <mgr_module.CommandResult object at
0x7fffcc990f50>, <mgr_module.CommandResult object at 0x7fffcc990bd0>,
<mgr_module.CommandResult object at 0x7ff
fcc990d90>, <mgr_module.CommandResult object at 0x7fffcc990ad0>,
<mgr_module.CommandResult object at 0x7fffcc990410>,
<mgr_module.CommandResult object at 0x7fffbed241d0>,
<mgr_module.CommandResult object at 0x7fff6a6caf90>, <mgr_module.Co
mmandResult object at 0x7fffbed242d0>, <mgr_module.CommandResult
object at 0x7fffbed24d90>, <mgr_module.CommandResult object at
0x7fffbed24d50>, <mgr_module.CommandResult object at 0x7fffbed24550>,
<mgr_module.CommandResult object at 0x7f
ffbed245d0>, <mgr_module.CommandResult object at 0x7fffbed24510>,
<mgr_module.CommandResult object at 0x7fffbed24690>,
<mgr_module.CommandResult object at 0x7fffbed24990>]
...
2020-05-01 08:43:16.733 7fffca074700 20 mgr[balancer] done
...
but these mappings do not show up in the osd dump. And a minute
later, the balancer tries again and comes up with a set of very
similar mappings (same from and to OSDs, slightly different PG
numbers) - and keeps going like that every minute without any progress
(the set of upmap entries stays the same, does not increase).
Andras
On 5/1/20 8:12 AM, Andras Pataki wrote:
> I'm wondering if anyone still sees issues with ceph-mgr using CPU and
> being unresponsive even in recent Nautilus releases. We upgraded our
> largest cluster from Mimic to Nautilus (14.2.8) recently - it has
> about 3500 OSDs. Now ceph-mgr is constantly at 100-200% CPU (1-2
> cores), and becomes unresponsive after a few minutes. The
> finisher-Mgr queue length grows (I've seen it at over 100k) - similar
> symptoms as seen with earlier Nautilus releases by many. This is what
> it looks like after an hour of running:
>
> "finisher-Mgr": {
> "queue_len": 66078,
> "complete_latency": {
> "avgcount": 21,
> "sum": 2098.408767721,
> "avgtime": 99.924227034
> }
> },
>
> We have a pretty vanilla manager config, only the balancer is enabled
> in upmap mode. Here are the enabled modules:
>
> "always_on_modules": [
> "balancer",
> "crash",
> "devicehealth",
> "orchestrator_cli",
> "progress",
> "rbd_support",
> "status",
> "volumes"
> ],
> "enabled_modules": [
> "restful"
> ],
>
> Any ideas or outstanding issues in this area?
>
> Andras
>
_______________________________________________
ceph-users mailing list -- ceph-users(a)ceph.io
To unsubscribe send an email to ceph-users-leave(a)ceph.io