Dynamic CPU Count changes and Oracle Behaviour

[Alzhy]

My DBA is complaining that CPUs should be fixed on an HP-UX environment. This according to her is due to the fact of the parallel servers does not adjust properly to CPU count. Is this true? And are there really any adverse issues with bumping up CPU count on a server w/o taking Oracle and the apps down?

[Chan 007]

Nelson,

Why not implement PRM?

007

[Alzhy]

Already are on certain systems -- but not with WLM and PSETS. Our use of the basic PRM is just ensure ORACLE has guaranteed CPU cycles during crunch time situations.

DBAs or Hybrids out there care to comment?


*Hybrids -- those potent DBAs who came from SysAdmin backgrounds... very very valuable these days.

[Steven E. Protter]

I'm a sysadmin who was an oracle dba in parellel. The dba part got to big so they hired one for me.

I have seen nothing to back up that conclusion. I'd like to see a document before I made changes.

http://metalink.oracle.com

SEP

[Geoff Wild]

How are you changing cpus? with ICAP/ICOD?

If so - then it doesn't matter - as those cpu's are "visible" to oracle - just that all can't be used - until you turn them on.

Rgds...Geoff

[Alzhy]

vPArs mi amigos... through the magic of Virtual Partitions.

[TwoProc]

As a long time hybrid, I'd say I can't see how any singular or parallel process would be adversely affected by suddenly getting MORE cpus.

I could see, I guess, a problem that if he the parallel server count were sized to a larger number (at the high points of cpu count), and then the query was required to run suddenly at a lower count.

If I hinted an a operation to run correctly with 24 cpus and I was really running with only 8 at some time - yeah that could conceivably be a problem.

[Bill Hassell]

As far as I know, the Oracle CPU count value just tells Oracle what is available. I don't think Oracle tries to verify the physical count so if you add CPUs, Oracle should not change it's multi-cpu behavior at all. The extra CPUs would be used by other apps. (this doesn't imply that Oracle uses specific CPUs, just the number of parallel server processes will not change)

[Alzhy]

Messr. Joubert,

I am aware of the Parallel server issue and we never actually size CPUs down but even if we did - what adverse effect would it have aside from the obvious performance impacts? My DBA is claiming that it results in a hang application.

What strikes me is our DBAs do not want to increase CPU count as well -- even if the system evidently needs it. The server does not run a Database load purely. AT crunch time - majority of the applications that consume CPU are not oracle fore/background processes but C/Cobol/Java executables that most of the time leave the Oracle apps w/ little or no chance at the CPUs.

On one machine where we have this same situation - I enabled PRM (for CPU) and guaranteed Oracle proccesses a set percentage of CPU time instead of the FSS (fair share scheduling) that HP-UX offers as default. And I am seeing realy that during 100% CPU utilization - Oracle gets at least its due CPU cycles to service the applications.

I also once was a "hybrid" but have been keeping mum about DBA affairs.

[Yogeeraj_1]

hi,

normally, the degree of parallelism in Oracle processes is not too much dependent on the underlying hardware -- the number of CPUs.

As you already know, you can perform queries in parallel to the order of many degrees with only one processor in the server.

I believe that the OS will automatically do the necessary adjustments and the Oracle Instance picking up resources as they are made available...

just some thoughts

kind regards
yogeeraj

[Tim Sanko]

From my experience It is the number of CPUs that has a bigger difference in performance. In ICOD the CPUS are visible to oracle but not active. We currently have 16 CPUS ACTIVE. WE have a faster 8 way but the 16 way handles our load better than the faster 8 way.

This may be due to the fact that we have 3000 users on the ERP server... it is both a DB and application server. Many users mean many processes. Our ERP likes many processors.

If you are doing it with partitioning I think it wouldn't be any worse.

Tim

[Ted Buis]

Could you use gang scheduling to keep the number of CPUs for Oracle fixed but allow them added CPUs to be used by other processes? It was originally designed for the HP Performance Technical Computing environment, but it might make sense here.