[shannong]

Does anyone have a good resource that compares Virtual CPUs to Fixed CPUs? I'm looking for performance differences and/or benchmarks?
Basically, I have a CPU intense application that I need to determine if I should put on Fixed CPUs or risk putting them on Virtual.
Thanks in advance!

[ross.mather]

The in-depth stuff is all in the Redbooks. It depends in large as to whether this CPU intensive application is in the only LPAR on a server or whether there are multiple LPARS on the same server.

A virtual CPU can only ever use upto one physical CPU so you cannot - even if you wanted to - define more Virtual CPUs than there are Physical Ones for an LPAR on a server.

The second thing to understand is how processor folding works related to desired CPU levels.

Can you articulate why you think using Virtual CPUs is a risk? I would always recommend them when ever possible (ie not running AIX 5.2).

[shannong]

The nature of the applicaiton is a 40TB datawarehouse. We were concerned that virutal CPUs will not have the capacity to handle the load. This application needs to handle 1TB of data load per hour. Do to the nature and importance of this applicaiton we decided to put them on fixed CPUs. Thanks for your input.

[john.samons]

Are you sure a CPU bottleneck will be your biggest problem? On what hardware are you going to run? Stand-alone or with other LPAR's? Available memory? Did you want to virtualize the CPU via an HMC (shared) or VIOS configuration?

I don't think there's a risk in using virtual CPU's, as in some cases you could actually benefit, i.e. from uncapped configurations etc. Have you checked if your application/database supports/favors SMT configurations?

[cdelgadop]

Hi

Hmmm. DatawareHouse behavior will need CPU power to run selects that does not use indexes to access data so along with best practices in configuring I/O devices you need your processors with all their power so i would configure dedicated CPUs for that LPAR.

Hope this helps

[shannong]

Thank you all for your input! I would have loved to use Virtual CPUs for this app especially since most of the heaviest processing is done in 4Q but I was not willing to bet my job on it. Which BTW, was my two options ;-)..."Give them fixed CPUs to make sure they have the resources they need, or give them virtual and if they don't meet their performance SLAs, turn in your badge"

[john.samons]

Well depending on your configuration Virtual CPU's can still work, however if you have Dedicated CPU's, go for it!

I normally set it to a shared CPU configuration, and set it up to use the CPU's as you would have in a dedicated configuration. So if you find you need to increase or decrease your CPU resources you can do it easily. If this partition starts early, i.e. first started partition, and your entitlement is 4.0 (out of 8.0), you will use an entitlement of 4 from 4 CPU's - so they should operate almost as dedicated. There is probably a very very small overhead, but not significant to really be worried about.

Dedicated resources, stand alone should always outperform virtualized & shared resources ~ or so they make us believe.......

[ross.mather]

OK to get the best performance in the config you are suggesting I would suggest the following:

Minimum and Desired CPU levels set to the number of CPUs you would normally allocated. Set the Maximum number of CPUs to the number of CPUs in the box. That way yoiu are guaranteed to have the relevant amount of CPU available.

The Virtual CPUs again set the Minimum to the number of CPUs you currently have and set the Desired and Maximum to the amount of CPU in the box.

Set this LPAR to be uncapped with a priority of 255 and the other LPARS on the box to 64 (or something lower).

This way your application is absoutely guaranteed to have your fixed level of CPU and it will also grab any other free CPU around in the box.

Of course I for one wouldb't blame you if you took job security of cool technology - which would probably actually make the application go faster......

Alternatively ask IBM or a Business Partner to come in and give you some advice that sometimes swings management over to your side.

[Mark Taylor]

i am not sure your terminology is correct .. do you mean Dedicated or Shared ? or Dedicated and Virtual ?

I think the rule of thumb for Dedicated, is that if your lpar is going to be using a high amount of cpu all the time, then you should use dedicated, and if you lpar is going to require cpu in peaks then shared would be the way to go allowing the free cycles to be used by other lpars .. shared requires cycles to dispatch to the correct virtual processor, so you would be wasting cycles if your cpu was going to be busy all the time.

Now, as for "how many virtual procs do I define" .. that totally depends on your application .. applications that make use of a multithreading "**MAY**" perform better with a higher number of virtual procs .. and something like a single threaded java proc is going to benefit from having access to the whole cpu, so a lower number of virtual procs.

its a suck it and see .. all applications are different and even thet way your company uses the application will be different from another implimentation.

Get a set of test cases that reflect your day to day business and benchmark them whilst changing the number of virtual procs ..

Rgds
Mark Taylor

[steevojb]

Hi,

I agree with Mark regarding your terminology and the 'suck it and see' approach. In addition, Ross point is a good suggestion regarding the weight value to ensure the data warehouse gets priority

Steve