[LSF/MM/BPF TOPIC] HA features for NVMe-oF

[Ballard, Curtis C (HPE Storage)]

Storage system vendors developing systems that present virtualized NVMe-oF
storage, such as storage arrays, are getting customer requests for HA features 
to support always-on infrastructures. Frequent requests are:
  * support of Active/Active where two physically separate storage systems 
    present the same NVMe namespace; and
  * support for non-disruptive migration at end-of-life for moving all storage
    volumes to new physical storage systems.

Since those storage systems are virtualized they are capable of managing the
relationships between the physical storage systems and the NVM subsystems
reported by those physical storage systems. There are methods that can be 
used to provide some of that functionality but there are complexities that 
make the models that have been attempted so far difficult to develop, test, 
and maintain. One of the methods most commonly discussed is usually
called "stretched subsystems" where the identity of a virtualized subsystem
is extended beyond a single physical storage system.

In this topic I'd like to present, in a multi-vendor presentation, on the customer
requirements and implementation challenges for HA NVMe-oF. I'd like to enlist
the help of Linux experts inside my company and in the community to help with an
overview of the Linux identity binding architecture.

Following the overview of the problem statement, we can discuss ideas for
enabling improved HA for NVMe-oF while innovating within the Linux NVMe-oF and
NVM Express architectures. The discussion should include considerations of
making changes in the NVM Express specifications. Participants that are prepared
to join this topic, if held, are also prepared to drive NVM Express changes in
support of any recommendations from this discussion.

I think it is very timely for this discussion as there is work happening in NVM
Express today on a technical proposal that touches on some of the same areas.
For those in NVM Express that technical proposal is known as TPAR4193. It would
be good to consider whether the preferred direction for HA on NVMe-oF is 
compatible with design decisions that are made for TP4193 if that work item
request is approved. This discussion may help refine thinking on the TP4193
topic as well.

Curtis Ballard
Strategist, Storage Technology
Hewlett Packard Enterprise