HA support of ALUA

[danisoto]

Hi,

StarWind SAN supports "ALUA" semantics in HA?

We need this for Multipath iSCSI with VSphere. Without it I don't know how to enable multipath for primary and set all paths to secondary as standby.

Please, I know how to enable multipathing in VSphere; and how to set fixed path policy for preferred primary. My question is about multiple active paths to "primary", and multiple standby paths to "secondary".

Thank you!

[lohelle]

This is something we want also. I hope this is possible as we have a very fast primary SAN (SSD) and slower SAS SAN as backup.

[anton (staff)]

From my understanding you just need to configure MRE policy (instead of Fixed or Round Robin) and that's all. StarWind really does not care where did particular request came from so pick up faster for read node as preferred path if you use unequal shoulders for cluster (writes should be the same as they touch cache memory and not media).

[danisoto]

From my understanding you just need to configure MRE policy ... StarWind really does not care where did particular request came from so pick up faster for read node as preferred path.

Hi Anton,

This is true with ONE path for each node! And this is right with BALANCED nodes.

But our implementation is with Unbalanced nodes: one node is fast (for writes and reads) and set as PRIMARY. Without ALUA semantics it's impossible to configure VSphere to use multiple paths to primary as preferred, and one or more as a standby for secondary.

In the Windows Inititator, MCS can be configured with the policy "Round Robin with a subset of paths", the one that I like to replicate in VSphere.

Can you provide this in the future?

[anton (staff)]

Do yo have any idea I could find a requirements for what should be done @ target side to support such a scenario? From what I've seen there were either configuration manuals how to enable ALUA on some array or papers describing why other array cannot work in AULA (b/c of deadlocks). So your feedback is appreciated here, we'll be happy to help.

From my understanding you just need to configure MRE policy ... StarWind really does not care where did particular request came from so pick up faster for read node as preferred path.

Hi Anton,

This is true with ONE path for each node! And this is right with BALANCED nodes.

But our implementation is with Unbalanced nodes: one node is fast (for writes and reads) and set as PRIMARY. Without ALUA semantics it's impossible to configure VSphere to use multiple paths to primary as preferred, and one or more as a standby for secondary.

In the Windows Inititator, MCS can be configured with the policy "Round Robin with a subset of paths", the one that I like to replicate in VSphere.

Can you provide this in the future?

[danisoto]

Do yo have any idea I could find a requirements for what should be done @ target side to support such a scenario? From what I've seen there were either configuration manuals how to enable ALUA on some array or papers describing why other array cannot work in AULA (b/c of deadlocks). So your feedback is appreciated here, we'll be happy to help.

The best documentation is the LIO target from Linux.
See http://linux-iscsi.org/wiki/Asymmetric_Logical_Unit_Assignment

The official documentation is the SPC-4 standard. But you may want to see the source code of the ALUA implementation in the LIO targat.

Regards.

[anton (staff)]

We're doing investigation about how soon we could have this one implemented in a proper way.

Do yo have any idea I could find a requirements for what should be done @ target side to support such a scenario? From what I've seen there were either configuration manuals how to enable ALUA on some array or papers describing why other array cannot work in AULA (b/c of deadlocks). So your feedback is appreciated here, we'll be happy to help.

The best documentation is the LIO target from Linux.
See http://linux-iscsi.org/wiki/Asymmetric_Logical_Unit_Assignment

The official documentation is the SPC-4 standard. But you may want to see the source code of the ALUA implementation in the LIO targat.

Regards.

[lohelle]

Testing 5.9 beta. Looks good!
ALUA support seems to work! Great work!!!
Se attachments. When using RR,and partner target selected as non-optimized node when creating the HA-device, only the paths to the optimized node are used. Failover and failback works!

[anton (staff)]

Good! Please give this one a heavier load to see would it melt or not We'll continue to stress it it the lab.

P.S. Super-fast RAM-based disk management is on the way as well Stay tuned!

[lohelle]

I will do some 3-node RAM-disk testing (use standalone ramdrive and create image-based normal HA-devices). Will run in VM's, but on 10G network

Attached is ALUA-test on a 3-node HA-setup. Second and third node set as non-optimized.
The only "problem" I see now is how to handle a 3-way setup when the third node is even slower than the second. When the optimized node go down, all the other paths are used (by AULA/VMware design I guess)

ALUA-support was just as cool as I hoped it would be!

[anton (staff)]

1) Make sure you use external app for RAM disk management. For now built-in RAM disk inside StarWind is SLOW (uses paged memory so spikes of load and delayes of I/O are possible). We're working on it now.

2) Don't test ALUA and 3-node config for now.

[lohelle]

2: Alua + 3-node does work for me there. Will do some more testing.
If I need 3 nodes, where ONE is very slow, I think async replication is the way to go..

Is the ALUA-support something that is complicated "inside Starwind", or is this just a "tag/value" that you add to the target so that the initiator see the target as a ALUA-array?

[Alex (staff)]

Is the ALUA-support something that is complicated "inside Starwind", or is this just a "tag/value" that you add to the target so that the initiator see the target as a ALUA-array?

It's additional functionality for processing ALUA-specific commands inside StarWind, not just simple tag on the target.