[gpfsug-discuss] maybe a silly question about "old school" gpfs
viccornell at gmail.com
Wed Nov 5 10:42:22 GMT 2014
> On 5 Nov 2014, at 10:33, Salvatore Di Nardo <sdinardo at ebi.ac.uk> wrote:
> I understand that my test its a bit particular because the client was also one of the servers.
> Usually clients don't have direct access to the storages, but still it made think, hot the things are supposed to work.
> For example i did another test with 3 dd's, one each server. All the servers was writing to all the luns.
> In other words a lun was accessed in parallel by 3 servers.
> Its that a problem, or gpfs manage properly the concurrency and avoid data corruption?
Its not a problem if you use locks. Remember the clients - even the ones running on the NSD servers are talking to the filesystem - not to the LUNS/NSDs directly.
It is the NSD processes that talk to the NSDs.
So loosely speaking it is as if all of the processes you are running were running on a single system with a local filesystem
So yes - gpfs is designed to manage the problems created by having a distributed, shared filesystem, and does a pretty good job IMHO.
> I'm asking because i was not expecting a server to write to an NSD he doesn't own, even if its locally available.
> I thought that the general availablity was for failover, not for parallel access.
Bear in mind that GPFS supports a number of access models, one of which is where all of the systems in the cluster have access to all of the disks.
So parallel access is most commonly used for failover, but that is not the limit of its capabilities.
> On 05/11/14 10:22, Vic Cornell wrote:
>> Hi Salvatore,
>> If you are doing the IO on the NSD server itself and it can see all of the NSDs it will use its "local” access to write to the LUNS.
>> You need some GPFS clients to see the workload spread across all of the NSD servers.
>>> On 5 Nov 2014, at 10:15, Salvatore Di Nardo <sdinardo at ebi.ac.uk <mailto:sdinardo at ebi.ac.uk>> wrote:
>>> Hello again,
>>> to understand better GPFS, recently i build up an test gpfs cluster using some old hardware that was going to be retired. THe storage was SAN devices, so instead to use native raids I went for the old school gpfs. the configuration is basically:
>>> 3x servers
>>> 3x san storages
>>> 2x san switches
>>> I did no zoning, so all the servers can see all the LUNs, but on nsd creation I gave each LUN a primary, secondary and third server. with the following rule:
>>> server2 server3
>>> storage2 server2 server3 server1
>>> storage3 server3 server1 server2
>>> looking at the mmcrnsd, it was my understanding that the primary server is the one that wrote on the NSD unless it fails, then the following server take the ownership of the lun.
>>> Now come the question:
>>> when i did from server 1 a dd surprisingly i discovered that server1 was writing to all the luns. the other 2 server was doing nothing. this behaviour surprises me because on GSS only the RG owner can write, so one server "ask" the other server to write to his own RG's.In fact on GSS can be seen a lot of ETH traffic and io/s on each server. While i understand that the situation it's different I'm puzzled about the fact that all the servers seems able to write to all the luns.
>>> SAN deviced usually should be connected to one server only, as paralled access could create data corruption. In environments where you connect a SAN to multiple servers ( example VMWARE cloud) its softeware task to avoid data overwriting between server ( and data corruption ).
>>> Honestly, what i was expecting is: server1 writing on his own luns, and data traffic ( ethernet) to the other 2 server , basically asking them to write on the other luns. I dont know if this behaviour its normal or not. I triied to find a documentation about that, but could not find any.
>>> Could somebody tell me if this "every server write to all the luns" its intended or not?
>>> Thanks in advance,
>>> gpfsug-discuss mailing list
>>> gpfsug-discuss at gpfsug.org <http://gpfsug.org/>
>>> http://gpfsug.org/mailman/listinfo/gpfsug-discuss <http://gpfsug.org/mailman/listinfo/gpfsug-discuss>
>> gpfsug-discuss mailing list
>> gpfsug-discuss at gpfsug.org
>> http://gpfsug.org/mailman/listinfo/gpfsug-discuss <http://gpfsug.org/mailman/listinfo/gpfsug-discuss>
> gpfsug-discuss mailing list
> gpfsug-discuss at gpfsug.org
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the gpfsug-discuss