OSDs don't actually delete files when two CephFS (or more) are in use

OSDs don't actually delete files when two CephFS (or more) are in use

[Geoffrey Hartz]

Hi!

I'm new to Ceph and I have a strange behavior with CephFS

Config is :

Ubuntu 12.04
Kernel 3.6.9
Ceph V0.55

2 OSD, 1 mon, 1 MDS, all on same host
2 clients, separate Hosts

Ceph.conf:

http://paste.ubuntu.com/1423712/

To mount the share I use : sudo ceph-fuse -m 192.168.80.139:6789 /mnt

When I create a file on one client, the other see the file, can be
downloaded etc.

But when I delete the file, both clients don't see the file anymore
BUT the file is still there on OSD (using space disk).

When I umount from BOTH clients, OSD are update and file is actually
delete (same behavior with mount -t ceph)

I'm missing something?

Thanks!

--
Geoffrey HARTZ

Re: OSDs don't actually delete files when two CephFS (or more) are in use

[Sam Lang]

On 12/10/2012 08:29 AM, Geoffrey Hartz wrote:
> Hi!
>
> I'm new to Ceph and I have a strange behavior with CephFS
>
> Config is :
>
> Ubuntu 12.04
> Kernel 3.6.9
> Ceph V0.55
>
> 2 OSD, 1 mon, 1 MDS, all on same host
> 2 clients, separate Hosts
>
> Ceph.conf:
>
> http://paste.ubuntu.com/1423712/
>
> To mount the share I use : sudo ceph-fuse -m 192.168.80.139:6789 /mnt
>
> When I create a file on one client, the other see the file, can be
> downloaded etc.
>
> But when I delete the file, both clients don't see the file anymore
> BUT the file is still there on OSD (using space disk).

Removing a file removes the directory entry (as you've seen), but the 
inode itself doesn't get removed until all references to it are dropped. 
  The clients may cache the capability for those inodes for a period of 
time, so you're not seeing the references drop until they get evicted 
from the cache.  Unmounting ensures that they get evicted from the 
client caches, so all references go to zero.

Also, removal of the underlying objects is done lazily, so you may not 
see the space get freed up right away.

-sam

>
> When I umount from BOTH clients, OSD are update and file is actually
> delete (same behavior with mount -t ceph)
>
> I'm missing something?
>
> Thanks!
>
> --
> Geoffrey HARTZ
> --
> To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>

Re: OSDs don't actually delete files when two CephFS (or more) are in use

[Geoffrey Hartz]

Hi.

I wait about more than 15 minutes. I was doing some benchmark when I
noticed that the space never go back to normal..

How can I disable this behavior? It's on Ceph side?

With one client, OSD are cleaned after few seconds, this sounds normal

2012/12/10 Sam Lang <sam.lang@inktank.com>:
> On 12/10/2012 08:29 AM, Geoffrey Hartz wrote:
>>
>> Hi!
>>
>> I'm new to Ceph and I have a strange behavior with CephFS
>>
>> Config is :
>>
>> Ubuntu 12.04
>> Kernel 3.6.9
>> Ceph V0.55
>>
>> 2 OSD, 1 mon, 1 MDS, all on same host
>> 2 clients, separate Hosts
>>
>> Ceph.conf:
>>
>> http://paste.ubuntu.com/1423712/
>>
>> To mount the share I use : sudo ceph-fuse -m 192.168.80.139:6789 /mnt
>>
>> When I create a file on one client, the other see the file, can be
>> downloaded etc.
>>
>> But when I delete the file, both clients don't see the file anymore
>> BUT the file is still there on OSD (using space disk).
>
>
> Removing a file removes the directory entry (as you've seen), but the inode
> itself doesn't get removed until all references to it are dropped.  The
> clients may cache the capability for those inodes for a period of time, so
> you're not seeing the references drop until they get evicted from the cache.
> Unmounting ensures that they get evicted from the client caches, so all
> references go to zero.
>
> Also, removal of the underlying objects is done lazily, so you may not see
> the space get freed up right away.
>
> -sam
>
>>
>> When I umount from BOTH clients, OSD are update and file is actually
>> delete (same behavior with mount -t ceph)
>>
>> I'm missing something?
>>
>> Thanks!
>>
>> --
>> Geoffrey HARTZ
>> --
>> To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
>> the body of a message to majordomo@vger.kernel.org
>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>>
>



-- 
Geoffrey HARTZ

Re: OSDs don't actually delete files when two CephFS (or more) are in use

[Sam Lang]

On 12/10/2012 10:34 AM, Geoffrey Hartz wrote:
> Hi.
>
> I wait about more than 15 minutes. I was doing some benchmark when I
> noticed that the space never go back to normal..
>
> How can I disable this behavior? It's on Ceph side?
>
> With one client, OSD are cleaned after few seconds, this sounds normal

Actually after discussing with Greg and Sage, it sounds like this is a 
bug.  The issue is that the client that didn't remove the file is 
caching the dentry indefinitely.  I've created a ticket to track the 
issue here:  http://tracker.newdream.net/issues/3601.

Unfortunately, I don't think there's a good workaround for the time 
being.  You can try to evict those cached dentries by creating/accessing 
a bunch of other files, but the default cache size on the client is 
16384, which is a lot of files to touch just to free up the space for 
those removed files. :-)

You can decrease the cache size with the config option client_cache_size:

[client]
	client cache size = 128

Then you only have to create/touch 128 files to evict the other files 
from the cache.  That's not ideal, because reducing the cache size will 
affect your overall performance, but if you know that you won't be 
accessing a lot of files anyway, its probably your best bet.

-sam

>
> 2012/12/10 Sam Lang <sam.lang@inktank.com>:
>> On 12/10/2012 08:29 AM, Geoffrey Hartz wrote:
>>>
>>> Hi!
>>>
>>> I'm new to Ceph and I have a strange behavior with CephFS
>>>
>>> Config is :
>>>
>>> Ubuntu 12.04
>>> Kernel 3.6.9
>>> Ceph V0.55
>>>
>>> 2 OSD, 1 mon, 1 MDS, all on same host
>>> 2 clients, separate Hosts
>>>
>>> Ceph.conf:
>>>
>>> http://paste.ubuntu.com/1423712/
>>>
>>> To mount the share I use : sudo ceph-fuse -m 192.168.80.139:6789 /mnt
>>>
>>> When I create a file on one client, the other see the file, can be
>>> downloaded etc.
>>>
>>> But when I delete the file, both clients don't see the file anymore
>>> BUT the file is still there on OSD (using space disk).
>>
>>
>> Removing a file removes the directory entry (as you've seen), but the inode
>> itself doesn't get removed until all references to it are dropped.  The
>> clients may cache the capability for those inodes for a period of time, so
>> you're not seeing the references drop until they get evicted from the cache.
>> Unmounting ensures that they get evicted from the client caches, so all
>> references go to zero.
>>
>> Also, removal of the underlying objects is done lazily, so you may not see
>> the space get freed up right away.
>>
>> -sam
>>
>>>
>>> When I umount from BOTH clients, OSD are update and file is actually
>>> delete (same behavior with mount -t ceph)
>>>
>>> I'm missing something?
>>>
>>> Thanks!
>>>
>>> --
>>> Geoffrey HARTZ
>>> --
>>> To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
>>> the body of a message to majordomo@vger.kernel.org
>>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>>>
>>
>
>
>

Re: OSDs don't actually delete files when two CephFS (or more) are in use

[Geoffrey Hartz]

Ok!

I'm not in production yet. Only testing. In about 1/2months I will use
in real (waiting for v0.56 bobtail)

Thanks for your consideration! really appreciate.


I have an other strange behavior between CephFS and kernel mount

With ceph-fuse command :
Client 1 create 2000 files and do "ls" => 1/2sec
Client 2 do "ls" => 1/2sec

With mount -t ceph ...
Client 1 create 2000 files and do "ls" => few ms
Client 2 do "ls" => 4-5sec

2012/12/10 Sam Lang <sam.lang@inktank.com>:
> On 12/10/2012 10:34 AM, Geoffrey Hartz wrote:
>>
>> Hi.
>>
>> I wait about more than 15 minutes. I was doing some benchmark when I
>> noticed that the space never go back to normal..
>>
>> How can I disable this behavior? It's on Ceph side?
>>
>> With one client, OSD are cleaned after few seconds, this sounds normal
>
>
> Actually after discussing with Greg and Sage, it sounds like this is a bug.
> The issue is that the client that didn't remove the file is caching the
> dentry indefinitely.  I've created a ticket to track the issue here:
> http://tracker.newdream.net/issues/3601.
>
> Unfortunately, I don't think there's a good workaround for the time being.
> You can try to evict those cached dentries by creating/accessing a bunch of
> other files, but the default cache size on the client is 16384, which is a
> lot of files to touch just to free up the space for those removed files. :-)
>
> You can decrease the cache size with the config option client_cache_size:
>
> [client]
>         client cache size = 128
>
> Then you only have to create/touch 128 files to evict the other files from
> the cache.  That's not ideal, because reducing the cache size will affect
> your overall performance, but if you know that you won't be accessing a lot
> of files anyway, its probably your best bet.
>
> -sam
>
>
>>
>> 2012/12/10 Sam Lang <sam.lang@inktank.com>:
>>>
>>> On 12/10/2012 08:29 AM, Geoffrey Hartz wrote:
>>>>
>>>>
>>>> Hi!
>>>>
>>>> I'm new to Ceph and I have a strange behavior with CephFS
>>>>
>>>> Config is :
>>>>
>>>> Ubuntu 12.04
>>>> Kernel 3.6.9
>>>> Ceph V0.55
>>>>
>>>> 2 OSD, 1 mon, 1 MDS, all on same host
>>>> 2 clients, separate Hosts
>>>>
>>>> Ceph.conf:
>>>>
>>>> http://paste.ubuntu.com/1423712/
>>>>
>>>> To mount the share I use : sudo ceph-fuse -m 192.168.80.139:6789 /mnt
>>>>
>>>> When I create a file on one client, the other see the file, can be
>>>> downloaded etc.
>>>>
>>>> But when I delete the file, both clients don't see the file anymore
>>>> BUT the file is still there on OSD (using space disk).
>>>
>>>
>>>
>>> Removing a file removes the directory entry (as you've seen), but the
>>> inode
>>> itself doesn't get removed until all references to it are dropped.  The
>>> clients may cache the capability for those inodes for a period of time,
>>> so
>>> you're not seeing the references drop until they get evicted from the
>>> cache.
>>> Unmounting ensures that they get evicted from the client caches, so all
>>> references go to zero.
>>>
>>> Also, removal of the underlying objects is done lazily, so you may not
>>> see
>>> the space get freed up right away.
>>>
>>> -sam
>>>
>>>>
>>>> When I umount from BOTH clients, OSD are update and file is actually
>>>> delete (same behavior with mount -t ceph)
>>>>
>>>> I'm missing something?
>>>>
>>>> Thanks!
>>>>
>>>> --
>>>> Geoffrey HARTZ
>>>> --
>>>> To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
>>>> the body of a message to majordomo@vger.kernel.org
>>>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>>>>
>>>
>>
>>
>>
>



-- 
Geoffrey HARTZ