MC ServiceGuard Calculation Model

[Darrel Louis]

Hi,

I'm looking for a calculation model or a formula on how to balance your system resources for the applications running on each node.
e.g
10 applications 4 nodes.
application1 CPU 20% Memory 30% etc.
application2 CPU 40% Memory 30% etc.
Application3 CPU 10% Memory 10% etc.
and so on.
Do you know if there is such a tool which can easily calculate what the system resources of a node will be when running application 1, 3 and 5 on it.

In my case we've 8 nodes and 44 packages.

Thanx in advanced

Darrel

[Bharat Katkar]

Hi Darrel,
I don't know how it can be used in Mc/Serviceguard Enviromment but if you have harwdare (HPUX servers) supporting Npars/Vpars (11iv2May2005 OS and relevant H/w) there is something called as "Instant Capacity on Demand i.e. Icod" which enable you to dynamically rotate resources as per the varying load.
If you are talking about existing setup then it would take total revamping which may not be possible.

Regards,

[Darrel Louis]

Hi Bharat,

This is not a Superdome or Vpar infra.
In fact what I'm looking for should fit in any infra.
I know it's not that easy to accomplish, but it should be possible.

Thanx

Darrel

[Patrick Wallek]

There really is not an easy way to develop such a tool. It is going to depend entirely on how the applications are used in your environment. Take Oracle for example -- One site may have a small Oracle DB that requires very few resources. Another site may have a Oracle DB with data that ranges into the terabytes that will require a whole lot more resources.

Your best bet is to look at your applications and your hardware and figure out the ratios. There will absolutely not be a cut-and-dried, one-size-fits-all answer to this question.

[Geoff Wild]

No tool - this is where you earn your money.

You have 8 nodes - 44 packages - do you have stand-by nodes?

Or are all nodes setup to run all packages? (that would be scary - because - if 7 of 8 node were to fail - highly unlikely - I doubt a single node could handle it all)...

You have to determine a way to split the apps up...maybe it would be best to segregate your nodes into 4 sets of 2? that would equate to 11 packages per set....

It's going to take some trial an error....

Hope this helps!

Rgds...Geoff

[Darrel Louis]

Geoff/Patrick,

It wasn't quite easy but I've done some investigation and made a simple calculation model only based on CPU.
We aren't experiencing any performance problems but we want to have the cluster better balanced, so that if one node fails it's secondary node should perform without any problems.
What I've done, used MWA & Openview Performance Manager:First I've checked if there is a server with a CPU, Memory or Disk botleneck problem, none but three servers CPU was almost 20% higher than the other 5.
I've exported(to excel) the CPU usage (GBL_CPU_TOTAL_UTIL; GBL_RUN_QUEUE and GBL_ACTIVE_PROC) per server(to check CPU usage per server) and because our parm (/var/opt/perf) is pre-defined per application, I could also export the CPU Utilization and Memory Utilization per application.

Now I can see what the expected CPU-usage and Memory Usage will be when the packages switch to another node, its not finished yet but its a start.
This is all manual intervention, it would have been easier if it can be done automaticly.

Darrel

[A. Clay Stephenson]

If you stop and think about this for a bit you will realize that this is an all but impossible task because very few applications are static in their use of resources. For example, an ERP application might be using 1% of resources at one point in time and 1 minute later be completely saturating the system. The best you can do is make reasonable assignments to split the load among the nodes during "normal" times but make sure that there are sufficient resources to at least allow all the packages to "limp along" during failover. I find that the most important criteria to meet is sufficient virtual memory because that is the one resource that will stop an application dead in its tracks. In UNIX, maxing out CPU, disk, and physical memory will still allow packages to run but insufficient virtual memory will produce the "Can't fork" or ENOMEM errors and processes will die (or never be spawned). Hopefully, you have enough physical memory that large amounts of swap are not necessary.

[Darrel Louis]

Clay,

You absolutely have a point, but at the moment the cluster isn't ideally configured, even its alternate nodes.
What I want to achieve is to have the nodes eqaully balanced with a 30% rule. In case of a Side failure the servers should cope with all packages running on one location.

We've enough VM available.

Thanx

Darrel

[Darrel Louis]

Gents thanx for the answers.