Timeslice parameter for 64-bit HP-UX v11.00 and Oracle 9.2.0

[Mott Given]

The timeslice parameter on a 64-bit HP-UX v11.00 system running Oracle 9.2.0 is set to 1.
What performance metrics would I evaluate to support my recommendation of changing timeslice to 10?

Mott Given

[Jeff Schussele]

Hi,

CHANGE IT IMMEDIATELY BACK TO 10!!!!!

Setting it to 1 will cause the system to do *TREMENDOUS* amounts of context switching which is pure housekeeping & nothing else.

Then throw that template away or do not follow that person's advice without verification again.

To verify check the run queue & especially the priority queue via glance or gpm.

My 2 cents,
Jeff

[Mott Given]

$ uptime
11:51am up 1 day, 4:29, 5 users, load average: 0.35, 0.38, 0.42

Glance shows the number of processes blocked on priority as 0. The machine is an HPV2250 with 12 processors.

$ top
Load averages: 0.33, 0.36, 0.41
391 processes: 332 sleeping, 14 running, 45 zombies


Mott

[A. Clay Stephenson]

Set it back to 10 and leave it alone. The best definition of timeslice that I have ever heard was given by Stephen Ciullo. He defined it as "the maximum number of ticks that a process ain't gonna get". If you set it to 1, the process is immediately switched out essentially without doing anything. The timeslice setting of 1 was an error in the tuned parameter set for databases and unfortunately has persisted and been copied to other systems. The smallest value that I would ever consider is 5 and in fact, pure database servers MIGHT do better with a larger than 10 timeslice setting but, in general, leave it at 10.

[Mott Given]

Apparently some of the Oracle install manuals recommend a timeslice value of 1.
I don't have any access to Metalink. Does anyone with access to Metalink or anything comparable know if Oracle has a special reason for recommending timeslice=1 instead of timeslice=10?

Mott

[Jeff Schussele]

Yep - it's called the sheep syndrome.
Somebody at HP mistakenly put a "suggested" kernel parameter template into SAM some years ago & people to this day still pass on this "nugget" of wisdom...errr...BS to be truthful because it really stinks.
I would have thought that Oracle, by now, would have seen the error of it's ways with regard to this.

My $0.02,
Jeff

[Kristopher March]

Can the timeslice parameter be changed dynamically? In otherwords, without a reboot of the OS.

[Kristopher March]

Disregard my last question. The answer is NO.

[John Poff · Accepted solution]

Hi,

I would suggest looking at the number of process context switches. Take a look at 'sar -w' and monitor the pswch/s column, which is the number of process context switches per second. If you can do a before and after measurement you might have the numbers to support your recommendation. I'm also guessing that you might see the system side of your CPU usage go down a little bit after changing timeslice to 10, as the box won't be so busy moving processes around.

We ran Oracle Financials on a V2500 box that had the timeslice set for 1 when we built the box using the Monolithic kernel template. It didn't kill the box, and when we set it up to 10 the box did run faster, but not so much that anybody got very excited. Remember, Oracle has lots of other ways to slow down your system.

JP

P.S. If they still won't let you change it, just do what I would do. Change it anyway, build the kernel, and then schedule yourself to accidentally trip across the power cord. ;)