NAS w/ multipath

[Tom Samplonius]

On Tue, 14 Feb 2006, Gary Palmer wrote:

> On Tue, Feb 07, 2006 at 12:22:33PM -0800, Tom Samplonius wrote:
>>   Now, in FreeBSD you could also multipath in the GEOM layer.  GEOM knows
>>   about devices going away, and knows how to handle that (ex. gmirror).
>> There is some support in GEOM for round-robin IO to two devices.  However,
>> phk has reported that the isp driver can hang forever on some timeouts, so
>> it might not be useful.  And I don't even know if GEOM round-robin is even
>> finished.
>
> Be very careful with that kind of work.  There are SAN units out there
> that do not share caches across multiple controllers.  To work around that
> limitation, they have one controller than can "own" the LUN at any one
> time, to ensure data consistency.  So if you write to a LUN through the
> controller which does not "own" the LUN, it actually transfers "ownership"
> of the LUN to the controller you sent the request through.  In at least

   Yes, trespass support.  Some controllers support auto-trespassing, but if the 
controllers do not have cache consistancy, you should probably make sure 
auto-trespass is disabled.  In manual trespassing, a specific SCSI command needs 
to be sent to activate the dormant LUN, which is generally vendor specific.


> one implimentation, the LUN vanishes from both controllers for a period
> measured in seconds while management of the LUN is handed off from one
> controller to the other. The vendor worked around this with special
> drivers that lived on the OS, which wasn't an option for us as they
> didn't have FreeBSD support.

   Yes, there are a lot of patches on the net to hack in various kinds of 
trespass support into Linux for various types of boxes.

> I suspect higher end devices (e.g. HDS and EMC Symmetrix units) this
> isn't a problem, but in mid range and lower end stuff I'd expect problems
> if the paths landed on separate controllers on the array.

   I don't think this is a problem with current mid-range stuff.  A mirrored 
write cache is considered a basic feature.  Not only does a mirrored write cache 
protect against controller cache consistancy, it also protects losing the 
contents of the write cache if a controller fails, which is generally a much 
bigger problem.

> Gary
>


Tom