[gpfsug-discuss] How to use RHEL 7 mdadm NVMe devices with Spectrum Scale 4.2.3.10?

[Lance Nakata]

We have a Dell R740xd with 24 x 1TB NVMe SSDs in the internal slots.  Since PERC
RAID cards don't see these devices, we are using mdadm software RAID to build
NSDs.  We took 12 NVMe SSDs and used mdadm to create a 10 + 1 + 1 hot spare RAID
5 stripe named /dev/md101.  We took the other 12 NVMe SSDs and created a similar
/dev/md102.

mmcrnsd worked without errors.  The problem is that Spectrum Scale does not see
the /dev/md10x devices as proper NSDs; the Device and Devtype columns are blank:

host2:~> sudo mmlsnsd -X

 Disk name    NSD volume ID      Device         Devtype  Node name                Remarks
---------------------------------------------------------------------------------------------------
 nsd0001      864FD12858A36E79   /dev/sdb       generic  host1.slac.stanford.edu server node
 nsd0002      864FD12858A36E7A   /dev/sdc       generic  host1.slac.stanford.edu server node
 nsd0021      864FD1285956B0A7   /dev/sdd       generic  host1.slac.stanford.edu server node
 nsd0251a     864FD1545BD0CCDF   /dev/dm-9      dmm      host2.slac.stanford.edu server node
 nsd0251b     864FD1545BD0CCE0   /dev/dm-11     dmm      host2.slac.stanford.edu server node
 nsd0252a     864FD1545BD0CCE1   /dev/dm-10     dmm      host2.slac.stanford.edu server node
 nsd0252b     864FD1545BD0CCE2   /dev/dm-8      dmm      host2.slac.stanford.edu server node
 nsd02nvme1   864FD1545BEC5D72   -              -        host2.slac.stanford.edu (not found) server node
 nsd02nvme2   864FD1545BEC5D73   -              -        host2.slac.stanford.edu (not found) server node

I know we can access the internal NVMe devices by their individual /dev/nvmeXX
paths, but non-ESS-based Spectrum Scale does not have built-in RAID
functionality.  Hence, the only option in that scenario is replication, which is
expensive and won't give us enough usable space.

Software Environment:
RHEL 7.6 with kernel 3.10.0-862.14.4.el7.x86_64
Spectrum Scale 4.2.3.10

Spectrum Scale Support has implied we can't use mdadm for NVMe devices.  Is that
really true?  Does anyone use an mdadm-based NVMe config?  If so, did you have
to do some kind of customization to get it working?

Thank you,

Lance Nakata
SLAC National Accelerator Laboratory