Load and relation to CPU

[Craig A. Sharp]

We have a 16 way dual core Dome. We consistantly see high CPU 90-100% as most of our processes are realtime. Our load's run around 1-1.5. Over 2.0 and we really see problems.
Now here is the question. What should this system be able to handle as far as load? Am I seeing a processor (process) bound system?
We have about 2000 telnet sessions at any one time..in addition, we run application and database on the same box.

[A. Clay Stephenson]

This is a very difficult question to answer given your unusual loads. Generally speaking, a system in which "most of our processes are realtime" is not one on which many time-shared processes (your 2000 telnet sessions) are also running -- or I should say not running because the RT processes always take precedence over TS processes. As long as there are runnable RT processes, no kernel process or TS process runs. You should really think about partitioning into RT and TS pieces so that the TS processes get much more predictable chunks of processor time.

[Volker Borowski]

Hi,

did you check out, the database is on latest patches ? For Oracle there have been some interimpatches that deal with excessive CPU usage in some cases (mind to remember it was on 9.2.0.5 and 9.2.0.6 at least for Solaris)

Volker

[Craig A. Sharp]

Yes, we run progress and we are current.

[Jeff Schussele]

Hi Craig,

"load" is simply a weighted average of the run queue size & is not an optimal measure to determine whether a system is CPU-bound.
Rup up glance or the GPM GUI & check the priority queue. That measures the avg number of processes that have been "bumped" off a CPU before it could complete it's task.
If you're seeing the priority is constantly > 2 then you can see if there are any reasons for the like memory starvation or slow disk or network I/O.

I highly recommend Stephen Ciullo's "HP-UX Performance Cookbook" - available here:

http://h21007.www2.hp.com/dspp/ddl/ddl_Download_File_TRX/1,1249280,00.pdf

HTH,
Jeff

[A. Clay Stephenson]

No amount of performance tuning is going to significantly help if "most of our processes are realtime". This will be especially true of all processes in the time-shared priority class. I am assuming that you are using "realtime" properly because that term has a very narrow definition in UNIX.

[Craig A. Sharp]

Thanks for all the info...