IIRC, the filesystem descriptor disk is only in the system pool… so as long as 
the system pool only has 3 FGs and they correspond to your 3 sites, then the 
filesystem survivability characteristics are straightforward.

I think that you technically could use two different FGs for the second pool 
and GPFS will still work as expected, but that just seems confusing to the 
humans.  We started off with one multisite stretch cluster like you describe 
and 10 years later we have around 100 stretch clusters.  Choosing a standard 
mapping between FG numbers and your sites can be a good way to reduce cognitive 
load on your team.

-Paul

From: gpfsug-discuss <[email protected]> On Behalf Of Luke 
Sudbery
Sent: Tuesday, March 18, 2025 16:18
To: gpfsug main discussion list <[email protected]>
Subject: Re: [gpfsug-discuss] Replicated cluster - failure groups


This message was sent by an external party.

But any pros/cons of 5 (2 failure groups per site + tiebreaker ) vs 3?

If we had 1 failure group per site, would we need to bring up all NSDs on that 
site (4x DSS – 8 actual servers) to guarantee bringing up the NSD with the desc 
replica disk?

Cheers,

Luke

--
Luke Sudbery
Principal Engineer (HPC and Storage).
Architecture, Infrastructure and Systems
Advanced Research Computing, IT Services
Room 132, Computer Centre G5, Elms Road

Please note I don’t work on Monday.

From: gpfsug-discuss 
<[email protected]<mailto:[email protected]>> 
On Behalf Of scale
Sent: 18 March 2025 17:21
To: gpfsug main discussion list 
<[email protected]<mailto:[email protected]>>
Subject: Re: [gpfsug-discuss] Replicated cluster - failure groups

CAUTION: This email originated from outside the organisation. Do not click 
links or open attachments unless you recognise the sender and know the content 
is safe.


There is no advantage to having more failure group than maximum number of 
replica supported by a file system plus 1 for tie breaker disks.    In a 
multiple site setup, you will want 1 failure group per site in order to ensure 
1 replica is placed at each site as GPFS will place replica using round-robin 
amount the failure groups.

From: gpfsug-discuss 
<[email protected]<mailto:[email protected]>> 
on behalf of Luke Sudbery 
<[email protected]<mailto:[email protected]>>
Date: Tuesday, March 18, 2025 at 9:28 AM
To: [email protected]<mailto:[email protected]> 
<[email protected]<mailto:[email protected]>>
Subject: [EXTERNAL] [gpfsug-discuss] Replicated cluster - failure groups
We are planning a replicated cluster. Due to a combination of purchasing 
cycles, floor loading and VAT-exemption status for half the equipment/data, 
this will be built over time using a total 8 Lenovo DSS building blocks. 2 main 
pools, in 2

We are planning a replicated cluster. Due to a combination of purchasing 
cycles, floor loading and VAT-exemption status for half the equipment/data, 
this will be built over time using a total 8 Lenovo DSS building blocks. 2 main 
pools, in 2 data centres, with 2 DSSG per pool, and a quorum/manager node with 
a local tie breaker disk in a 3rd physical location.

My main question is about failure groups - so far, with 2 DSS and 1 tiebreaker, 
we would have had 1 failure group per DSS and 1 for the tie breaker disk, 
giving us a total of 3. But if we did that now we would have 9 failure groups 
in 1 filesystem, which is more than the maximum number of replicas of the file 
system descriptor and not desirable, as I understand it.

So we could have either:

  *   1 FG per physical site, and assign all 4 DSS per site to 1 FG, and a 3rd 
to the tiebreaker
  *   1 FG per pool per site, with 2 DSS in each FG. This makes sense as the 
pairs of DSSG will both always need to be up for all the data in the pool to be 
accessible.

The second option would give us 5 failure groups, but what would be the 
advantage and disadvantages of more failure groups?

Many thanks,

Luke

--
Luke Sudbery
Principal Engineer (HPC and Storage).
Architecture, Infrastructure and Systems
Advanced Research Computing, IT Services
Room 132, Computer Centre G5, Elms Road

Please note I don’t work on Monday.

_______________________________________________
gpfsug-discuss mailing list
gpfsug-discuss at gpfsug.org
http://gpfsug.org/mailman/listinfo/gpfsug-discuss_gpfsug.org

Reply via email to