Re: Ceph, LIO, VMWARE anyone?

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Oh yeah, I am not completely sure (have not tested myself), but if you
were doing a setup where you were not using a clustering app like
windows/redhat clustering that uses PRs, did not use vmfs and were
instead accessing the disks exported by LIO/TGT directly in the vm
(either using the guest's iscsi client or as a raw esx device), and were
not using ESX clustering, then you might be safe doing active/passive or
active/active with no modifications needed other than some scripts to
distribute the setup info across LIO/TGT nodes.

Were any of you trying this type of setup when you were describing your
results? If so, were you running oracle or something like that? Just
wondering.


On 01/27/2015 08:58 PM, Mike Christie wrote:
> I do not know about perf, but here is some info on what is safe and
> general info.
> 
> - If you are not using VAAI then it will use older style RESERVE/RELEASE
> commands only.
> 
> If you are using VAAI ATS, and doing active/active then you need
> something, like the lock/sync talked about in the slides/hammer doc,
> that would coordinate multiple ATS/COMPARE_AND_WRITEs from executing at
> the same time on the same sectors. You probably do not ever see problems
> today, because it seems ESX normally does this command for only one
> sector and I do not think there are multiple commands for the same
> sectors in flight normally.
> 
> For active/passive, ATS is simple since you only have the one LIO/TGT
> node executing commands at a time, so the locking is done locally using
> a normal old mutex.
> 
> - tgt and LIO both support SCSI-3 persistent reservations. This is not
> really needed for ESX vmfs though since it uses ATS or older
> RESERVE/RELEASE. If you were using a cluster app like windows
> clustering, red hat cluster, etc in ESX or in normal non vm use, then
> you need something extra to support SCSI-3 PRs in both active/active or
> active/passive.
> 
> For AA, you need something like described in that doc/video.
> 
> For AP, you would need to copy over the PR state from one node to the
> other when failing over/back across nodes. For LIO this is in /var/target.
> 
> Depending on how you do AP (what ALUA states you use if you do ALUA),
> you might also need to always distribute the PR info if you are doing
> windows clustering. Windows wants to see a consistent view of the PR
> info from all ports if you do something like ALUA active-optimized and
> standby states for active/passive.
> 
> - I do not completely understand the comment about using LIO as a
> backend for tgt. You would either use tgt or LIO to export a rbd device.
> Not both at the same time like using LIO for some sort of tgt backend.
> Maybe people meant using the "RBD backend" instead of "LIO backend"
> 
> - There are some other setup complications that you can see here
> http://comments.gmane.org/gmane.linux.scsi.target.devel/7044
> if you are using ALUA. I think tgt does not support ALUA, but LIO does.
> 
> 
> On 01/23/2015 04:25 PM, Zoltan Arnold Nagy wrote:
>> Correct me if I'm wrong, but tgt doesn't have full SCSI-3 persistence
>> support when _not_ using the LIO
>> backend for it, right?
>>
>> AFAIK you can either run tgt with it's own iSCSI implementation or you
>> can use tgt to manage your LIO targets.
>>
>> I assume when you're running tgt with the rbd backend code you're
>> skipping all the in-kernel LIO parts (in which case
>> the RedHat patches won't help a bit), and you won't have proper
>> active-active support, since the initiators
>> have no way to synchronize state (and more importantly, no way to 
>> synchronize write caching! [I can think
>> of some really ugly hacks to get around that, tho...]).
>>
>> On 01/23/2015 05:46 PM, Jake Young wrote:
>>> Thanks for the feedback Nick and Zoltan,
>>>
>>> I have been seeing periodic kernel panics when I used LIO.  It was
>>> either due to LIO or the kernel rbd mapping.  I have seen this on
>>> Ubuntu precise with kernel 3.14.14 and again in Ubunty trusty with the
>>> utopic kernel (currently 3.16.0-28).  Ironically, this is the primary
>>> reason I started exploring a redundancy solution for my iSCSI proxy
>>> node.  So, yes, these crashes have nothing to do with running the
>>> Active/Active setup.
>>>
>>> I am moving my entire setup from LIO to rbd enabled tgt, which I've
>>> found to be much more stable and gives equivalent performance.
>>>
>>> I've been testing active/active LIO since July of 2014 with VMWare and
>>> I've never seen any vmfs corruption.  I am now convinced (thanks Nick)
>>> that it is possible.  The reason I have not seen any corruption may
>>> have to do with how VMWare happens to be configured.
>>>
>>> Originally, I had made a point to use round robin path selection in
>>> the VMware hosts; but as I did performance testing, I found that it
>>> actually didn't help performance.  When the host switches iSCSI
>>> targets there is a short "spin up time" for LIO to get to 100% IO
>>> capability.  Since round robin switches targets every 30 seconds (60
>>> seconds? I forget), this seemed to be significant.  A secondary goal
>>> for me was to end up with a config that required minimal tuning from
>>> VMWare and the target software; so the obvious choice is to leave
>>> VMWare's path selection at the default which is Fixed and picks the
>>> first target in ASCII-betical order.  That means I am actually
>>> functioning in Active/Passive mode.
>>>
>>> Jake
>>>
>>>
>>>
>>>
>>> On Fri, Jan 23, 2015 at 8:46 AM, Zoltan Arnold Nagy
>>> <zoltan@xxxxxxxxxxxxxxxxxx <mailto:zoltan@xxxxxxxxxxxxxxxxxx>> wrote:
>>>
>>>     Just to chime in: it will look fine, feel fine, but underneath
>>>     it's quite easy to get VMFS corruption. Happened in our tests.
>>>     Also if you're running LIO, from time to time expect a kernel
>>>     panic (haven't tried with the latest upstream, as I've been using
>>>     Ubuntu 14.04 on my "export" hosts for the test, so might have
>>>     improved...).
>>>
>>>     As of now I would not recommend this setup without being aware of
>>>     the risks involved.
>>>
>>>     There have been a few upstream patches getting the LIO code in
>>>     better cluster-aware shape, but no idea if they have been merged
>>>     yet. I know RedHat has a guy on this.
>>>
>>>     On 01/21/2015 02:40 PM, Nick Fisk wrote:
>>>>
>>>>     Hi Jake,
>>>>
>>>>      
>>>>
>>>>     Thanks for this, I have been going through this and have a pretty
>>>>     good idea on what you are doing now, however I maybe missing
>>>>     something looking through your scripts, but I’m still not quite
>>>>     understanding how you are managing to make sure locking is
>>>>     happening with the ESXi ATS SCSI command.
>>>>
>>>>      
>>>>
>>>>     From this slide
>>>>
>>>>      
>>>>
>>>>     http://xo4t.mjt.lu/link/xo4t/gzyhtx3/1/_9gJVMUrSdvzGXYaZfCkVA/aHR0cHM6Ly93aWtpLmNlcGguY29tL0BhcGkvZGVraS9maWxlcy8zOC9oYW1tZXItY2VwaC1kZXZlbC1zdW1taXQtc2NzaS10YXJnZXQtY2x1c3RlcmluZy5wZGY  
>>>>     (Page 8)
>>>>
>>>>      
>>>>
>>>>     It seems to indicate that for a true active/active setup the two
>>>>     targets need to be aware of each other and exchange locking
>>>>     information for it to work reliably, I’ve also watched the video
>>>>     from the Ceph developer summit where this is discussed and it
>>>>     seems that Ceph+Kernel need changes to allow this locking to be
>>>>     pushed back to the RBD layer so it can be shared, from what I can
>>>>     see browsing through the Linux Git Repo, these patches haven’t
>>>>     made the mainline kernel yet.
>>>>
>>>>      
>>>>
>>>>     Can you shed any light on this? As tempting as having
>>>>     active/active is, I’m wary about using the configuration until I
>>>>     understand how the locking is working and if fringe cases
>>>>     involving multiple ESXi hosts writing to the same LUN on
>>>>     different targets could spell disaster.
>>>>
>>>>      
>>>>
>>>>     Many thanks,
>>>>
>>>>     Nick
>>>>
>>>>      
>>>>
>>>>     *From:*Jake Young [mailto:jak3kaj@xxxxxxxxx]
>>>>     *Sent:* 14 January 2015 16:54
>>>>
>>>>
>>>>     *To:* Nick Fisk
>>>>     *Cc:* Giuseppe Civitella; ceph-users
>>>>     *Subject:* Re:  Ceph, LIO, VMWARE anyone?
>>>>
>>>>      
>>>>
>>>>     Yes, it's active/active and I found that VMWare can switch from
>>>>     path to path with no issues or service impact.
>>>>
>>>>      
>>>>
>>>>       
>>>>
>>>>     I posted some config files here: github.com/jak3kaj/misc
>>>>     <http://xo4t.mjt.lu/link/xo4t/gzyhtx3/2/_P2HWj3RxQZC1v5DQ_206Q/aHR0cDovL2dpdGh1Yi5jb20vamFrM2thai9taXNj>
>>>>
>>>>      
>>>>
>>>>     One set is from my LIO nodes, both the primary and secondary
>>>>     configs so you can see what I needed to make unique.  The other
>>>>     set (targets.conf) are from my tgt nodes.  They are both 4 LUN
>>>>     configs.
>>>>
>>>>      
>>>>
>>>>     Like I said in my previous email, there is no performance
>>>>     difference between LIO and tgt.  The only service I'm running on
>>>>     these nodes is a single iscsi target instance (either LIO or tgt).
>>>>
>>>>      
>>>>
>>>>     Jake
>>>>
>>>>      
>>>>
>>>>     On Wed, Jan 14, 2015 at 8:41 AM, Nick Fisk <nick@xxxxxxxxxx
>>>>     <mailto:nick@xxxxxxxxxx>> wrote:
>>>>
>>>>         Hi Jake,
>>>>
>>>>          
>>>>
>>>>         I can’t remember the exact details, but it was something to
>>>>         do with a potential problem when using the pacemaker resource
>>>>         agents. I think it was to do with a potential hanging issue
>>>>         when one LUN on a shared target failed and then it tried to
>>>>         kill all the other LUNS to fail the target over to another
>>>>         host. This then leaves the TCM part of LIO locking the RBD
>>>>         which also can’t fail over.
>>>>
>>>>          
>>>>
>>>>         That said I did try multiple LUNS on one target as a test and
>>>>         didn’t experience any problems.
>>>>
>>>>          
>>>>
>>>>         I’m interested in the way you have your setup configured
>>>>         though. Are you saying you effectively have an active/active
>>>>         configuration with a path going to either host, or are you
>>>>         failing the iSCSI IP between hosts? If it’s the former, have
>>>>         you had any problems with scsi locking/reservations…etc
>>>>         between the two targets?
>>>>
>>>>          
>>>>
>>>>         I can see the advantage to that configuration as you
>>>>         reduce/eliminate a lot of the troubles I have had with
>>>>         resources failing over.
>>>>
>>>>          
>>>>
>>>>         Nick
>>>>
>>>>          
>>>>
>>>>         *From:*Jake Young [mailto:jak3kaj@xxxxxxxxx
>>>>         <mailto:jak3kaj@xxxxxxxxx>]
>>>>         *Sent:* 14 January 2015 12:50
>>>>         *To:* Nick Fisk
>>>>         *Cc:* Giuseppe Civitella; ceph-users
>>>>         *Subject:* Re:  Ceph, LIO, VMWARE anyone?
>>>>
>>>>          
>>>>
>>>>         Nick,
>>>>
>>>>          
>>>>
>>>>         Where did you read that having more than 1 LUN per target
>>>>         causes stability problems?
>>>>
>>>>          
>>>>
>>>>         I am running 4 LUNs per target. 
>>>>
>>>>          
>>>>
>>>>         For HA I'm running two linux iscsi target servers that map
>>>>         the same 4 rbd images. The two targets have the same serial
>>>>         numbers, T10 address, etc.  I copy the primary's config to
>>>>         the backup and change IPs. This way VMWare thinks they are
>>>>         different target IPs on the same host. This has worked very
>>>>         well for me. 
>>>>
>>>>          
>>>>
>>>>         One suggestion I have is to try using rbd enabled tgt. The
>>>>         performance is equivalent to LIO, but I found it is much
>>>>         better at recovering from a cluster outage. I've had LIO lock
>>>>         up the kernel or simply not recognize that the rbd images are
>>>>         available; where tgt will eventually present the rbd images
>>>>         again. 
>>>>
>>>>          
>>>>
>>>>         I have been slowly adding servers and am expanding my test
>>>>         setup to a production setup (nice thing about ceph). I now
>>>>         have 6 OSD hosts with 7 disks on each. I'm using the LSI
>>>>         Nytro cache raid controller, so I don't have a separate
>>>>         journal and have 40Gb networking. I plan to add another 6 OSD
>>>>         hosts in another rack in the next 6 months (and then another
>>>>         6 next year). I'm doing 3x replication, so I want to end up
>>>>         with 3 racks. 
>>>>
>>>>          
>>>>
>>>>         Jake
>>>>
>>>>         On Wednesday, January 14, 2015, Nick Fisk <nick@xxxxxxxxxx
>>>>         <mailto:nick@xxxxxxxxxx>> wrote:
>>>>
>>>>             Hi Giuseppe,
>>>>
>>>>              
>>>>
>>>>             I am working on something very similar at the moment. I
>>>>             currently have it working on some test hardware but seems
>>>>             to be working reasonably well.
>>>>
>>>>              
>>>>
>>>>             I say reasonably as I have had a few instability’s but
>>>>             these are on the HA side, the LIO and RBD side of things
>>>>             have been rock solid so far. The main problems I have had
>>>>             seem to be around recovering from failure with resources
>>>>             ending up in a unmanaged state. I’m not currently using
>>>>             fencing so this may be part of the cause.
>>>>
>>>>              
>>>>
>>>>             As a brief description of my configuration.
>>>>
>>>>              
>>>>
>>>>             4 Hosts each having 2 OSD’s also running the monitor role
>>>>
>>>>             3 additional host in a HA cluster which act as iSCSI
>>>>             proxy nodes.
>>>>
>>>>              
>>>>
>>>>             I’m using the IP, RBD, iSCSITarget and iSCSILUN resource
>>>>             agents to provide HA iSCSI LUN which maps back to a RBD.
>>>>             All the agents for each RBD are in a group so they follow
>>>>             each other between hosts.
>>>>
>>>>              
>>>>
>>>>             I’m using 1 LUN per target as I read somewhere there are
>>>>             stability problems using more than 1 LUN per target.
>>>>
>>>>              
>>>>
>>>>             Performance seems ok, I can get about 1.2k random IO’s
>>>>             out the iSCSI LUN. These seems to be about right for the
>>>>             Ceph cluster size, so I don’t think the LIO part is
>>>>             causing any significant overhead.
>>>>
>>>>              
>>>>
>>>>             We should be getting our production hardware shortly
>>>>             which wil have 40 OSD’s with journals and a SSD caching
>>>>             tier, so within the next month or so I will have a better
>>>>             idea of running it in a production environment and the
>>>>             performance of the system.
>>>>
>>>>              
>>>>
>>>>             Hope that helps, if you have any questions, please let me
>>>>             know.
>>>>
>>>>              
>>>>
>>>>             Nick
>>>>
>>>>              
>>>>
>>>>             *From:*ceph-users
>>>>             [mailto:ceph-users-bounces@xxxxxxxxxxxxxx] *On Behalf Of
>>>>             *Giuseppe Civitella
>>>>             *Sent:* 13 January 2015 11:23
>>>>             *To:* ceph-users
>>>>             *Subject:*  Ceph, LIO, VMWARE anyone?
>>>>
>>>>              
>>>>
>>>>             Hi all,
>>>>
>>>>              
>>>>
>>>>             I'm working on a lab setup regarding Ceph serving rbd
>>>>             images as ISCSI datastores to VMWARE via a LIO box. Is
>>>>             there someone that already did something similar wanting
>>>>             to share some knowledge? Any production deployments? What
>>>>             about LIO's HA and luns' performances?
>>>>
>>>>              
>>>>
>>>>             Thanks 
>>>>
>>>>             Giuseppe
>>>>
>>>>
>>>>
>>>>      
>>>>
>>>>
>>>>
>>>>
>>>>     _______________________________________________
>>>>     ceph-users mailing list
>>>>     ceph-users@xxxxxxxxxxxxxx <mailto: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





[Index of Archives]     [Information on CEPH]     [Linux Filesystem Development]     [Ceph Development]     [Ceph Large]     [Ceph Dev]     [Linux USB Development]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]     [xfs]


  Powered by Linux