Help with CPU bottle neck

[Marie-Josee Cormier]

Very two or three days, we hit a cpu bottle neck on our system. We very new to this OS (11i) as we were on OpenVMS. We're currently migrating.

We are now in the test phase, with some users doing some tests, but once in a while, we run into this problems. I read a bit on the subject, and refering to HP docs, this is not a network/memory/or Disk I/O issue. Glance tells me that it's 100% sure it's a CPU Botleneck issue.

The problem is, when we hit this state, top can't seem to help me, like maybe my system is too busy to produre an accurate top report.

I attached a report one of HP script created, and hope somebody can help me with this!

Thanks!!!

[A. Clay Stephenson]

The very first thing you should do is reduce the default dbc_max_pct from 50% to a much smaller value (10%?). Since I have no idea how much memory you have, I can't tell you a good value but in general, even on boxes with very large amounts of memory, you want no more than 800-1600MB of cache. I actually prefer a fixed cache by setting bufpages to a non-zero value. After that value is adjusted, I suspect that your metrics will be more meaningful. It would also help if you describe your application, amount of memory, speed and number of processors. I note that you are doing a huge amount of logical i/o amnd little physical i/o. This suggests inefficient algorithms which read the same blocks over and over.

[B. Hulst]

Hi,

Change the nbuf value to a non-zero value.
It makes the system use a fix disk buffer size.
nbuf=0 is a variable disk buffer size.
Your system seems to have lots of NFS and SMB connections...

Regards,
Bob

[A. Clay Stephenson]

Wrong. Leave nbuf=0, the default behavior is fine. If you want to implement a fixed size buffer cache, set bufpages to the number of 4K cache pages you want (e.g bufpages=204800 for a fixed 800MB cache).

[Marie-Josee Cormier]

Thanks for all of your replies. And yes, that's true that I forgot two important thing. I have 2gigs of ram on this system. And my CPU is an Itanium 750 MHZ. So with this number, I think the dbc_max_pct at 50% should be alright, since it gives me a 1000MB cache.

I'll try doing as you said and set the nbuf to a 1000MB cache just to see how well my system goes after this change.

It's hard for me to think this thing only will fix this, but i'll try. I should be able to reboot my system in the dinner time, so i'll stay in touch with you.

Thanks!

[Bharat Katkar]

Marie,

May be this cookbook help you.

Regards,

[A. Clay Stephenson]

A box with only 2GB of memory is considered very small these days and in this case a 1000MB buffer cache is much too large. I would set dbc_mac_pct to no more than 12 (~240MB) on a box with your amount of memory.

[Alzhy]

On those occassions that you're machine is simply overwhelmed - I think you are just simply feeding it more than it can chew. On VMS/Mainframe you can throttle processes but on a plain jane UNIX systems -- its practically a free for all. Try using the command:

UNIX95=1 ps -efxH

and observe how your processes are spawned during your heaviest use of your server. Those cache/kernel adjustments will cetainly help but it will all boil down to proper scheduling of processes (or throttling) and just sheer resource issues. (BTW, we use PRM on our machines where we want to tame executions of processes).

And with just 2 x 750 Mhz CPUs - I think you will as well need a faster more capable machine.

[Zeev Schultz]

Some observations - noticed that load splited close to equally between user and kernel space,however user applications were dominant and got more load usually. I also saw sometimes high number of fork system call,meaning many child proccesses were created fast.I would first go with general performance suggestions like those in the pdf posted above.