[dm-devel] [RFC] [PATCH] add serial keyword to the weightedpath prioritizer

[Hannes Reinecke]

On 08/01/2016 10:42 AM, Christophe Varoqui wrote:
>
>
> On Mon, Aug 1, 2016 at 9:49 AM, Hannes Reinecke <hare at suse.de
> <mailto:hare at suse.de>> wrote:
>
>     On 07/31/2016 09:26 PM, Christophe Varoqui wrote:
>
>         Ben, Hannes,
>
>         Can you review this patch, adding a new 'serial' keyword to the
>         weightedpath prioritizer.
>
>         I compile-tested it only, as I have no testing environment at
>         hand at
>         the moment.
>
>         I commited it in a separate 'weightedpath-serial' branch for now.
>
>         http://git.opensvc.com/?p=multipath-tools/.git;a=commitdiff;h=4dd16d99281104fc3504ad73626894a5c3702fb3
>
>         Thanks,
>         Christophe Varoqui
>         OpenSVC
>
>     Well.
>     In general, sure, fine, I don't have any issues with that.
>     If the customer wants to diddle with his array that way...
>
>     The more general problem I'm seeing is that our current two-layered
>     priority setup (path groups with distinct priorities and paths
>     within them) might not be leading to issues with larger and more
>     complex scenarios.
>
>     ATM we already have the problem that clustered scenarios like this:
>
>     Storage node 1(active):
>       Path 1 (optimal):
>         LUN 1, LUN2
>       Path 2 (non-optimal):
>         LUN 1, LUN2
>
>     Storage node 2(passive):
>       Path 1(optimal):
>          LUN 1, LUN2
>       Path 2(non-optimal):
>          LUN 1, LUN2
>
>     can not be represented properly with multipath tools.
>     We are forced to either
>     a) set 'storage node 2' to 'failed', which would kill
>        any cluster instance accessing only 'storage node 2'
>     or
>     b) map all priorities from 'storage node 2' to '0',
>        thereby losing all priority information
>
>     Things become even more convoluted if both storage nodes are in fact
>     accessible, or if someone would be using different transports.
>
> Would something like "prio alua+weightedpath" produce correct priorities
> for the path grouping ? where priorities reported by alua would be added
> to those reported by weighted path. That syntax extension would reduce
> the need to develop more complex prioritizers.
>
Hmm.
Allowing stacked prioritizers is a nice idea.
But then we need to impose some preference here; if we do not set any 
restrictions on the value of the prioritizers we end up with a jumble of 
(essentially unreadable) priorities.
EG if your weightedpath returns values of '5' or '0' they'll be readily 
obscured by alua information, which uses '5' for the non-optimized path.

So if we were to got that route we need to restrict the values of the 
prioritizers to eg 256, and shift the stacked prioritizer values ontop 
of each other.
EG with a stacked 'prio_alua+weightedpath' we should end up with a 
priority of 0xAAWW.
With that we can allow up to 4 levels of stacking (or 8 if we extend 
that to 64 bits), and still keep source-level compability with the 
original code.
We could even reduce the permissive values for the prioritzers even 
more; 16 is enough even for ALUA, and that would leave us with enough 
room of 1024 possible stacking levels :-)

But in general I like the idea.

Cheers,

Hannes
-- 
Dr. Hannes Reinecke		   Teamlead Storage & Networking
hare at suse.de			               +49 911 74053 688
SUSE LINUX GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: F. Imendörffer, J. Smithard, J. Guild, D. Upmanyu, G. Norton
HRB 21284 (AG Nürnberg)