Friday, October 22, 2010

AOS and clustering


There seems to be quite some vagueness when it comes to Ax and AOS clustering.
A few statements to (hopefully) clear things up a bit:

Single AOS
The simplest configuration is an Ax implementation with one single AOS server.
Extra AOS servers can be added in scenarios where you want to increase availability or scale up the Ax environment to handle a larger number of users/connections.
Several possibilities from that point on:

AOS side by side
By default multiple AOS's in a single Ax environment will exist side by side. They do know about each others existence, sync their caches, try not to hamper each other too much,  …
A user connects to a specific AOS as defined in his client configuration. In case there is more than one AOS defined in the configuration the first available will be used.
Adding all available AOS's to one client config file would probably be a smart thing to do. If you only add one node of the cluster to your config file, and this one node is down, you won't be able to connect. If you add all - let's say - 4 nodes, the client config will connect to the first available.
Unless you want to control they way users connect and have a specific department or region connect to a specific AOS of course. I'm thinking latency, bandwidth, ...

Clustered AOS
You can create a load balancing cluster (administration - setup - cluster configuration) and link AOS instances to it. Again two options:

Cluster without load balancer
The naming here is a bit tricky, even if there is no dedicated load balancer set up in a cluster, there will still be load balancing: all AOS's in a cluster implicitly act as load balancers
A user connects to a any of the AOS's as defined in his client configuration. He cannot predict which AOS he will eventually host his session.
The only criteria maintained to balance the load is the number of clients (including business connectors, worker threads, ...) connected to an AOS. If a cluster has two nodes: AOS A and AOS B and A has 3 client connections and B has 4. The next connection will most likely be hosted by A, even if the connecting user's client config only has AOS B in the list of available AOS instances. Most likely, because if AOS A does not respond (in a fashionable time), it will try to connect to B.
In this same example it is possible the load (CPU, memory usage, ...) created by those 3 connections on AOS A is much higher than the load generated by the 4 connections on AOS B. Still the next connection will be hosted by A.
Therefore the client config should have all AOS's listed. You're not sure which AOS will host your session, but you do want all entry point available.

Cluster with load balancer
If you set up an AOS instance to be a load balancer (checkbox 'load balancer' in server configuration) you dedicate  this instance to spread incoming user connection among the non-load balancing AOS's. Ideally the one  with the lowest number of active connections in the cluster will host the first incoming connection. That's all the dedicated load balancer will do for you.
Frankly I don't see the point of setting up a dedicated load balancing server. Maybe when you're dealing with a huge number of concurrent users, logging on and off all the time and you don't want a 'real' AOS to waste time handling those? On the other side: a dedicated load balancer does not consume an AOS license. So it's actually free.

The client config would only contain the dedicated load balancing AOS instance(s), because that's the only valid entry point here. But then again, if your load balancer is down …
- Is it possible to directly connect to a non-load balancer AOS in an AOS cluster with dedicated load balancer? Yes it is.
- Does a non-load balancer AOS in an AOS cluster with a dedicated load balancer also balances load?
Yes it does.
So even if you've got a dedicated load balancer … setting the non-dedicated AOS's instances from the cluster in the client config may be a satisfying plan.
You can go all fancy an set up difference clusters on one Ax environment, for example to split the online and batch processing. Most likely you want client config files per AOS cluster to control who's connecting to which cluster.

A word on failover might also be in place. An AOS cluster does not mean failover capabilities. It just indicates two or more AOS instances are grouped. That's about it. No automatic moving of sessions from AOS A to AOS B if server A crashes, users will lose their connection to the AOS, manual restart of Ax client by user is required, …
All bad news? No ma'am: an AOS cluster makes sure the load is equally spread over the different nodes. You can add or remove nodes to the cluster online. But in fact … that's about it … and besides the load spreading, a cluster doesn't do more than a well considered config file does.

Other things you might not yet know regarding AOS's:
- Each active not-dedicated loadbalancing AOS instance requires one AOS license. You can install as many AOS instances as you wish, but the number of active instances is limited by the license. You cannot have more active AOS instances than your license allows on the same Ax environment.
- In the online users form you only get to see the SPID's of the users that are on the same AOS as you are
- You can prevent users to connect to an AOS using the 'reject new clients' button on the 'server instances' tab of the online users form. Hit 'accept new clients' to undo this.

No comments:

Post a Comment