[ClusterLabs] Antw: Resource placement strategy and utilization AND resource location preference

[Ulrich Windl]

>>> "Vallevand, Mark K" <Mark.Vallevand at UNISYS.com> schrieb am 27.10.2015 um 22:24
in Nachricht
<37343dddcd2d454baaea374e80d73363 at US-EXCH13-5.na.uis.unisys.com>:
> How do the resource placement strategy and utilization AND resource location 
> preference relate?
> 
> I mean, is it one or the other?  Or both somehow?

I think it's all AND (if you use -inf): So if one fails, the resource fails.

> 
> If I set a resource location preference, how will that affect placement 
> strategy like balanced?  Vice versa.
> 
> Here's the problem I'm looking at.
> I have a large number of resources that have very different utilization 
> values.  Say 1-10.  All my nodes have the same utilization values.  I want the 
> placement to be balanced.  That works nicely.  Consider what happens when a 

That's default AFAIK.

> node fails and then rejoins the cluster.  The balanced placement moves the 
> resources when the node fails and again when it rejoins.  It's not good to 

Depending on stickiness of resources.

> have resources move.  Setting a resource-stickiness helps when the node fails. 

You got it!

>  Rebalancing seems to be sane.  But, when the node rejoins, the resources 
> stick where they are and the rejoining node carries no load.  If I don't do 

You'll have to decide what you want: Should resources move, or shouldn't they?


> any resource placement strategy at all and consider each resource to be 
> equal, I can set resource location preferences so that resources move when 
> the node fails and return to it when it rejoins.

You could also write a script that checks the status and issues manual migration commands to the cluster to do what you want.

> I want it all.  :-)
> I want the resources to be placed with balanced regard to utilization.

Utilization does not balance,, but limit IMHO.

> I want only the resources on a failed node to be reallocated to remaining 
> nodes (with balanced utilization as much as possible).

Then use high stickiness.

> I want those resources to return to the node when it rejoins.  (Or a subset 
> of them if that balances better.)

Then you'll have to use a low stickiness.

> 
> I could ignore placement strategy and script up resource location 
> preferences that mimic a balanced load.  But, I'd rather let clustering do 
> it.

Honestly: Why do you care if one node has little work, while others can handle the load? Modern hardware can save significant energy when being idle.

> 
> Any ideas would be very welcome.

No more ideas ;-)

> 
> Regards.
> Mark K Vallevand   Mark.Vallevand at Unisys.com<mailto:Mark.Vallevand at Unisys.com>
> Never try and teach a pig to sing: it's a waste of time, and it annoys the 
> pig.
> THIS COMMUNICATION MAY CONTAIN CONFIDENTIAL AND/OR OTHERWISE PROPRIETARY 
> MATERIAL and is thus for use only by the intended recipient. If you received 
> this in error, please contact the sender and delete the e-mail and its 
> attachments from all computers.