On Wed, Oct 25, 2017 at 2:32 PM, Bryan Stillwell <bstillwell@xxxxxxxxxxx> wrote:
> That helps a little bit, but overall the process would take years at this
> rate:
>
>
>
> # for i in {1..3600}; do ceph df -f json-pretty |grep -A7 '".rgw.buckets"'
> |grep objects; sleep 60; done
>
> "objects": 1660775838
>
> "objects": 1660775733
>
> "objects": 1660775548
>
> "objects": 1660774825
>
> "objects": 1660774790
>
> "objects": 1660774735
>
>
>
> This is on a hammer cluster. Would upgrading to Jewel or Luminous speed up
> this process at all?
>
I'm not sure it's going to help much, although the omap performance
might improve there. The big problem is that the omaps are just too
big, so that every operation on them take considerable time. I think
the best way forward there is to take a list of all the rados objects
that need to be removed from the gc omaps, and then get rid of the gc
objects themselves (newer ones will be created, this time using the
new configurable). Then remove the objects manually (and concurrently)
using the rados command line tool.
The one problem I see here is that even just removal of objects with
large omaps can affect the availability of the osds that hold these
objects. I discussed that now with Josh, and we think the best way to
deal with that is not to remove the gc objects immediatly, but to
rename the gc pool, and create a new one (with appropriate number of
pgs). This way new gc entries will now go into the new gc pool (with
higher number of gc shards), and you don't need to remove the old gc
objects (thus no osd availability problem). Then you can start
trimming the old gc objects (on the old renamed pool) by using the
rados command. It'll take a very very long time, but the process
should pick up speed slowly, as the objects shrink.
Yehuda
>
>
> Bryan
>
>
>
> From: Yehuda Sadeh-Weinraub <yehuda@xxxxxxxxxx>
> Date: Wednesday, October 25, 2017 at 11:32 AM
> To: Bryan Stillwell <bstillwell@xxxxxxxxxxx>
> Cc: David Turner <drakonstein@xxxxxxxxx>, Ben Hines <bhines@xxxxxxxxx>,
> "ceph-users@xxxxxxxxxxxxxx" <ceph-users@xxxxxxxxxxxxxx>
>
>
> Subject: Re: Speeding up garbage collection in RGW
>
>
>
> Some of the options there won't do much for you as they'll only affect
>
> newer object removals. I think the default number of gc objects is
>
> just inadequate for your needs. You can try manually running
>
> 'radosgw-admin gc process' concurrently (for the start 2 or 3
>
> processes), see if it makes any dent there. I think one of the problem
>
> is that the gc omaps grew so much that operations on them are too
>
> slow.
>
>
>
> Yehuda
>
>
>
> On Wed, Oct 25, 2017 at 9:05 AM, Bryan Stillwell <bstillwell@xxxxxxxxxxx>
> wrote:
>
> We tried various options like the one's Ben mentioned to speed up the
> garbage collection process and were unsuccessful. Luckily, we had the
> ability to create a new cluster and move all the data that wasn't part of
> the POC which created our problem.
>
>
>
> One of the things we ran into was the .rgw.gc pool became too large to
> handle drive failures without taking down the cluster. We eventually had to
> move that pool to SSDs just to get the cluster healthy. It was not obvious
> it was getting large though, because this is what it looked like in the
> 'ceph df' output:
>
>
>
> NAME ID USED %USED MAX AVAIL OBJECTS
>
> .rgw.gc 17 0 0 235G
> 2647
>
>
>
> However, if you look at the SSDs we used (repurposed journal SSDs to get out
> of the disaster) in 'ceph osd df' you can see quite a bit of data is being
> used:
>
>
>
> 410 0.20000 1.00000 181G 23090M 158G 12.44 0.18
>
> 411 0.20000 1.00000 181G 29105M 152G 15.68 0.22
>
> 412 0.20000 1.00000 181G 110G 72223M 61.08 0.86
>
> 413 0.20000 1.00000 181G 42964M 139G 23.15 0.33
>
> 414 0.20000 1.00000 181G 33530M 148G 18.07 0.26
>
> 415 0.20000 1.00000 181G 38420M 143G 20.70 0.29
>
> 416 0.20000 1.00000 181G 92215M 93355M 49.69 0.70
>
> 417 0.20000 1.00000 181G 64730M 118G 34.88 0.49
>
> 418 0.20000 1.00000 181G 61353M 121G 33.06 0.47
>
> 419 0.20000 1.00000 181G 77168M 105G 41.58 0.59
>
>
>
> That's ~560G of omap data for the .rgw.gc pool that isn't being reported in
> 'ceph df'.
>
>
>
> Right now the cluster is still around while we wait to verify the new
> cluster isn't missing anything. So if there is anything the RGW developers
> would like to try on it to speed up the gc process, we should be able to do
> that.
>
>
>
> Bryan
>
>
>
> From: ceph-users <ceph-users-bounces@xxxxxxxxxxxxxx> on behalf of David
> Turner <drakonstein@xxxxxxxxx>
>
> Date: Tuesday, October 24, 2017 at 4:07 PM
>
> To: Ben Hines <bhines@xxxxxxxxx>
>
> Cc: "ceph-users@xxxxxxxxxxxxxx" <ceph-users@xxxxxxxxxxxxxx>
>
> Subject: Re: Speeding up garbage collection in RGW
>
>
>
> Thank you so much for chiming in, Ben.
>
>
>
> Can you explain what each setting value means? I believe I understand min
> wait, that's just how long to wait before allowing the object to be cleaned
> up. gc max objs is how many will be cleaned up during each period? gc
> processor period is how often it will kick off gc to clean things up? And
> gc processor max time is the longest the process can run after the period
> starts? Is that about right for that? I read somewhere saying that prime
> numbers are optimal for gc max objs. Do you know why that is? I notice
> you're using one there. What is lc max objs? I couldn't find a reference
> for that setting.
>
>
>
> Additionally, do you know if the radosgw-admin gc list is ever cleaned up,
> or is it an ever growing list? I got up to 3.6 Billion objects in the list
> before I killed the gc list command.
>
>
>
> On Tue, Oct 24, 2017 at 4:47 PM Ben Hines <bhines@xxxxxxxxx> wrote:
>
> I agree the settings are rather confusing. We also have many millions of
> objects and had this trouble, so i set these rather aggressive gc settings
> on our cluster which result in gc almost always running. We also use
> lifecycles to expire objects.
>
>
>
> rgw lifecycle work time = 00:01-23:59
>
> rgw gc max objs = 2647
>
> rgw lc max objs = 2647
>
> rgw gc obj min wait = 300
>
> rgw gc processor period = 600
>
> rgw gc processor max time = 600
>
>
>
>
>
> -Ben
>
>
>
> On Tue, Oct 24, 2017 at 9:25 AM, David Turner <drakonstein@xxxxxxxxx> wrote:
>
> As I'm looking into this more and more, I'm realizing how big of a problem
> garbage collection has been in our clusters. The biggest cluster has over 1
> billion objects in its gc list (the command is still running, it just
> recently passed by the 1B mark). Does anyone have any guidance on what to
> do to optimize the gc settings to hopefully/eventually catch up on this as
> well as stay caught up once we are? I'm not expecting an overnight fix, but
> something that could feasibly be caught up within 6 months would be
> wonderful.
>
>
>
> On Mon, Oct 23, 2017 at 11:18 AM David Turner <drakonstein@xxxxxxxxx> wrote:
>
> We recently deleted a bucket that was no longer needed that had 400TB of
> data in it to help as our cluster is getting quite full. That should free
> up about 30% of our cluster used space, but in the last week we haven't seen
> nearly a fraction of that free up yet. I left the cluster with this running
> over the weekend to try to help `radosgw-admin --rgw-realm=local gc
> process`, but it didn't seem to put a dent into it. Our regular ingestion
> is faster than how fast the garbage collection is cleaning stuff up, but our
> regular ingestion is less than 2% growth at it's maximum.
>
>
>
> As of yesterday our gc list was over 350GB when dumped into a file (I had to
> stop it as the disk I was redirecting the output to was almost full). In
> the future I will use the --bypass-gc option to avoid the cleanup, but is
> there a way to speed up the gc once you're in this position? There were
> about 8M objects that were deleted from this bucket. I've come across a few
> references to the rgw-gc settings in the config, but nothing that explained
> the times well enough for me to feel comfortable doing anything with them.
>
>
>
> On Tue, Jul 25, 2017 at 4:01 PM Bryan Stillwell <bstillwell@xxxxxxxxxxx>
> wrote:
>
> Excellent, thank you! It does exist in 0.94.10! :)
>
>
>
> Bryan
>
>
>
> From: Pavan Rallabhandi <PRallabhandi@xxxxxxxxxxxxxxx>
>
> Date: Tuesday, July 25, 2017 at 11:21 AM
>
>
>
> To: Bryan Stillwell <bstillwell@xxxxxxxxxxx>, "ceph-users@xxxxxxxxxxxxxx"
> <ceph-users@xxxxxxxxxxxxxx>
>
> Subject: Re: Speeding up garbage collection in RGW
>
>
>
> I’ve just realized that the option is present in Hammer (0.94.10) as well,
> you should try that.
>
>
>
> From: Bryan Stillwell <bstillwell@xxxxxxxxxxx>
>
> Date: Tuesday, 25 July 2017 at 9:45 PM
>
> To: Pavan Rallabhandi <PRallabhandi@xxxxxxxxxxxxxxx>,
> "ceph-users@xxxxxxxxxxxxxx" <ceph-users@xxxxxxxxxxxxxx>
>
> Subject: EXT: Re: Speeding up garbage collection in RGW
>
>
>
> Unfortunately, we're on hammer still (0.94.10). That option looks like it
> would work better, so maybe it's time to move the upgrade up in the
> schedule.
>
>
>
> I've been playing with the various gc options and I haven't seen any
> speedups like we would need to remove them in a reasonable amount of time.
>
>
>
> Thanks,
>
> Bryan
>
>
>
> From: Pavan Rallabhandi <PRallabhandi@xxxxxxxxxxxxxxx>
>
> Date: Tuesday, July 25, 2017 at 3:00 AM
>
> To: Bryan Stillwell <bstillwell@xxxxxxxxxxx>, "ceph-users@xxxxxxxxxxxxxx"
> <ceph-users@xxxxxxxxxxxxxx>
>
> Subject: Re: Speeding up garbage collection in RGW
>
>
>
> If your Ceph version is >=Jewel, you can try the `--bypass-gc` option in
> radosgw-admin, which would remove the tails objects as well without marking
> them to be GCed.
>
>
>
> Thanks,
>
>
>
> On 25/07/17, 1:34 AM, "ceph-users on behalf of Bryan Stillwell"
> <ceph-users-bounces@xxxxxxxxxxxxxx on behalf of bstillwell@xxxxxxxxxxx>
> wrote:
>
>
>
> I'm in the process of cleaning up a test that an internal customer did
> on our production cluster that produced over a billion objects spread across
> 6000 buckets. So far I've been removing the buckets like this:
>
>
>
> printf %s\\n bucket{1..6000} | xargs -I{} -n 1 -P 32 radosgw-admin
> bucket rm --bucket={} --purge-objects
>
>
>
> However, the disk usage doesn't seem to be getting reduced at the same
> rate the objects are being removed. From what I can tell a large number of
> the objects are waiting for garbage collection.
>
>
>
> When I first read the docs it sounded like the garbage collector would
> only remove 32 objects every hour, but after looking through the logs I'm
> seeing about 55,000 objects removed every hour. That's about 1.3 million a
> day, so at this rate it'll take a couple years to clean up the rest! For
> comparison, the purge-objects command above is removing (but not GC'ing)
> about 30 million objects a day, so a much more manageable 33 days to finish.
>
>
>
> I've done some digging and it appears like I should be changing these
> configuration options:
>
>
>
> rgw gc max objs (default: 32)
>
> rgw gc obj min wait (default: 7200)
>
> rgw gc processor max time (default: 3600)
>
> rgw gc processor period (default: 3600)
>
>
>
> A few questions I have though are:
>
>
>
> Should 'rgw gc processor max time' and 'rgw gc processor period' always
> be set to the same value?
>
>
>
> Which would be better, increasing 'rgw gc max objs' to something like
> 1024, or reducing the 'rgw gc processor' times to something like 60 seconds?
>
>
>
> Any other guidance on the best way to adjust these values?
>
>
>
> Thanks,
>
> Bryan
>
>
>
>
>
> _______________________________________________
>
> ceph-users mailing list
>
> ceph-users@xxxxxxxxxxxxxx
>
> http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
>
>
>
>
>
>
>
> _______________________________________________
>
> ceph-users mailing list
>
> ceph-users@xxxxxxxxxxxxxx
>
> http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
>
>
>
> _______________________________________________
>
> ceph-users mailing list
>
> ceph-users@xxxxxxxxxxxxxx
>
> http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
>
>
>
>
>
> _______________________________________________
>
> ceph-users mailing list
>
> ceph-users@xxxxxxxxxxxxxx
>
> http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
>
>
_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
