Seems like a message bus would be nice. Each opener of an RBD could subscribe for messages on the bus for that RBD. Anytime the map is modified a message could be put on the bus to update the others. That opens up a whole other can of worms though.
Robert LeBlanc
Sent from a mobile device please excuse any typos.
On Jan 6, 2015 5:35 PM, "Josh Durgin" <josh.durgin@xxxxxxxxxxx> wrote:
On 01/06/2015 04:19 PM, Robert LeBlanc wrote:
The bitmap certainly sounds like it would help shortcut a lot of code
that Xiaoxi mentions. Is the idea that the client caches the bitmap
for the RBD so it know which OSDs to contact (thus saving a round trip
to the OSD), or only for the OSD to know which objects exist on it's
disk?
It's purely at the rbd level, so librbd caches it and maintains its
consistency. The idea is that since it's kept consistent, librbd can do
things like delete exactly the objects that exist without any
extra communication with the osds. Many things that were
O(size of image) become O(written objects in image).
The only restriction is that keeping the object map consistent requires
a single writer, so this does not work for the rare case of e.g. ocfs2
on top of rbd, where there are multiple clients writing to the same
rbd image at once.
Josh
On Tue, Jan 6, 2015 at 4:19 PM, Josh Durgin <josh.durgin@xxxxxxxxxxx> wrote:
On 01/06/2015 10:24 AM, Robert LeBlanc wrote:
Can't this be done in parallel? If the OSD doesn't have an object then
it is a noop and should be pretty quick. The number of outstanding
operations can be limited to 100 or a 1000 which would provide a
balance between speed and performance impact if there is data to be
trimmed. I'm not a big fan of a "--skip-trimming" option as there is
the potential to leave some orphan objects that may not be cleaned up
correctly.
Yeah, a --skip-trimming option seems a bit dangerous. This trimming
actually is parallelized (10 ops at once by default, changeable via
--rbd-concurrent-management-ops) since dumpling.
What will really help without being dangerous is keeping a map of
object existence [1]. This will avoid any unnecessary trimming
automatically, and it should be possible to add to existing images.
It should be in hammer.
Josh
[1] https://github.com/ceph/ceph/pull/2700
On Tue, Jan 6, 2015 at 8:09 AM, Jake Young <jak3kaj@xxxxxxxxx> wrote:
On Monday, January 5, 2015, Chen, Xiaoxi <xiaoxi.chen@xxxxxxxxx> wrote:
When you shrinking the RBD, most of the time was spent on
librbd/internal.cc::trim_image(), in this function, client will iterator
all
unnecessary objects(no matter whether it exists) and delete them.
So in this case, when Edwin shrinking his RBD from 650PB to 650GB,
there are[ (650PB * 1024GB/PB -650GB) * 1024MB/GB ] / 4MB/Object =
170,227,200 Objects need to be deleted.That will definitely take a long
time
since rbd client need to send a delete request to OSD, OSD need to find
out
the object context and delete(or doesn’t exist at all). The time needed
to
trim an image is ratio to the size needed to trim.
make another image of the correct size and copy your VM's file system to
the new image, then delete the old one will NOT help in general, just
because delete the old volume will take exactly the same time as
shrinking ,
they both need to call trim_image().
The solution in my mind may be we can provide a “—skip-triming” flag to
skip the trimming. When the administrator absolutely sure there is no
written have taken place in the shrinking area(that means there is no
object
created in these area), they can use this flag to skip the time
consuming
trimming.
How do you think?
That sounds like a good solution. Like doing "undo grow image"
From: Jake Young [mailto:jak3kaj@xxxxxxxxx]
Sent: Monday, January 5, 2015 9:45 PM
To: Chen, Xiaoxi
Cc: Edwin Peer; ceph-users@xxxxxxxxxxxxxx
Subject: Re: rbd resize (shrink) taking forever and a day
On Sunday, January 4, 2015, Chen, Xiaoxi <xiaoxi.chen@xxxxxxxxx> wrote:
You could use rbd info <volume_name> to see the block_name_prefix, the
object name consist like <block_name_prefix>.<sequence_number>, so for
example, rb.0.ff53.3d1b58ba.00000000e6ad should be the <e6ad>th object
of
the volume with block_name_prefix rb.0.ff53.3d1b58ba.
$ rbd info huge
rbd image 'huge':
size 1024 TB in 268435456 objects
order 22 (4096 kB objects)
block_name_prefix: rb.0.8a14.2ae8944a
format: 1
-----Original Message-----
From: ceph-users [mailto:ceph-users-bounces@lists.ceph.com] On Behalf Of
Edwin Peer
Sent: Monday, January 5, 2015 3:55 AM
To: ceph-users@xxxxxxxxxxxxxx
Subject: Re: rbd resize (shrink) taking forever and a day
Also, which rbd objects are of interest?
<snip>
ganymede ~ # rados -p client-disk-img0 ls | wc -l
1672636
</snip>
And, all of them have cryptic names like:
rb.0.ff53.3d1b58ba.00000000e6ad
rb.0.6d386.1d545c4d.000000011461
rb.0.50703.3804823e.000000001c28
rb.0.1073e.3d1b58ba.00000000b715
rb.0.1d76.2ae8944a.00000000022d
which seem to bear no resemblance to the actual image names that the rbd
command line tools understands?
Regards,
Edwin Peer
On 01/04/2015 08:48 PM, Jake Young wrote:
On Sunday, January 4, 2015, Dyweni - Ceph-Users
<6EXbab4FYk8H@xxxxxxxxxx <mailto:6EXbab4FYk8H@dyweni.com>> wrote:
Hi,
If its the only think in your pool, you could try deleting the
pool instead.
I found that to be faster in my testing; I had created 500TB when
I meant to create 500GB.
Note for the Devs: I would be nice if rbd create/resize would
accept sizes with units (i.e. MB GB TB PB, etc).
On 2015-01-04 08:45, Edwin Peer wrote:
Hi there,
I did something stupid while growing an rbd image. I
accidentally
mistook the units of the resize command for bytes instead of
megabytes
and grew an rbd image to 650PB instead of 650GB. This all
happened
instantaneously enough, but trying to rectify the mistake is
not going
nearly as well.
<snip>
ganymede ~ # rbd resize --size 665600 --allow-shrink
client-disk-img0/vol-x318644f-0
Resizing image: 1% complete...
</snip>
It took a couple days before it started showing 1% complete
and has
been stuck on 1% for a couple more. At this rate, I should be
able to
shrink the image back to the intended size in about 2016.
Any ideas?
Regards,
Edwin Peer
_______________________________________________
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
You can just delete the rbd header. See Sebastien's excellent blog:
http://www.sebastien-han.fr/blog/2013/12/12/rbd-image-bigger-than-your
-ceph-cluster/
Jake
_______________________________________________
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
Sorry, I misunderstood.
The simplest approach to me is to make another image of the correct size
and copy your VM's file system to the new image, then delete the old
one.
The safest thing to do would be to mount the new file system from the VM
and do all the formatting / copying from there (the same way you'd move
a
physical server's root disk to a new physical disk)
I would not attempt to hack the rbd header. You open yourself up to some
unforeseen problems.
Unless one of the ceph developers can comment there is a safe way to
shrink an image, assuming we know that the file system has not grown
since
growing the disk.
Jake
_______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
