Load Balanced Clustering with Unequal Servers?

[#1]

Hi Everyone,

Can Load Balanced Clustering be made to work well with unequal servers?

I've got a happy situation where a client has outgrown their primary Production server, and has the greenlight to acquire a new, 4x more powerful server.  They will retain the old server, primarily because it's a sunk cost and still functional.  Can load balanced clustering be made to work in an uneven environment like this, where perhaps 75% of the apps are hosted on the new server and 25% are on the old server?  (Both are 64-bit, so this is not mixing environments.)

In the past, it appeared that load balancing a cluster assumed that each server was equal or nearly equal in capabilities, but the QlikView Server v.11.20 manual says the following, with reopens the question:

"The load balancing is determined by an internal ranking system based on the amount of memory available and the CPU use." p.114

Has anyone had experience with this?

Thanks,

DJ

[danielrozental · Accepted solution]

The paragraph you're quoting relates to a QlikView Publisher Cluster, not a QlikView Server one. You should check page 105.

QlikView AccessPoint supports three load balancing strategies:

- Random (default setting): A round robin type strategy ideal for most users, since the session is

distributed across all nodes in the cluster.

- Loaded document: Used when sessions for the same document are to be routed to the same server.

This strategy is designed for deployments where there are more documents than a single node in the

cluster can handle. AccessPoint makes the decision based on if the document is already loaded and

on the amount of RAM available on the server.

- CPU with RAM overload (only available in QlikView 11): Allows QlikView Web Server (QVWS) to

route traffic based on two factors, (1) RAM and (2) CPU use. The node is chosen using the following

criteria:

       - If RAM is readily available (low) on all available nodes, choose the node with the lowest CPU use.

       - If RAM is moderately used on all available nodes, choose the node with the most RAM available.

Another idea might be to separate server and publisher and use the old server just for reloads.

[Bill_Britt]

David,

You can cluster unlike servers, but there could be issues. You would need to be on a version Qlikview that will allow you to preload the documents on a certain server and you would have to preload all the documents, to make sure a big one does get load on the smaller server.

Version 11 will do this..

Bill

[dde]

Also just to add to what Bill said there is another caveat too. 

Remember that any QVS settings are cluster wide, for example Working Set Limits and CPU Affinity.  So for example if you take Working Set you may have to pick the lowest common denominator setting that will work on all nodes in the cluster.

[#5]

Thanks, everyone.  This is great information.

I imagine we will use a non-clustered method then, maybe having the older server perform reloads, while the new server concentrates on serving the end users.

Thanks,

DJ

[StefanBackstrand]

David; I personally like your approach better. Tiering your environment is a better option if performance is sufficient on each machine for their respective role.