tune kernel

[Alianto]

Hi All,

I have a D370 server which is running on hpux 11.0. Since the memory resources is my contraints, I am quite often tune its kernel whenever there are problem related with memory. For example, currently i have 11 instances of Oracle database. Before the recent kernel tuning, i am only able to startup max 10 instances or even 9 instances, but that situation did not guarantee the system will perform well. sometime i encountered very slow response, such as log in, issue 'bdf' command, etc.

But after the recent kernel tuning, the problem was disappeared. i can even startup 11 instances and without encountering slow response of actions.

Here the following example of kernel parameters that I changed,
nflocks from 256 to 384
nproc,nfile was depended of maxusers. and maxusers was tuned from 120 to 150.

that's only the example.

Now my question is
1. How far are we able to allow us to change the number ? for example, can i change nflocks parameter from 256 to 1024 or even 2048 ? or maxusers is tuned to 1024 or 8192
2. Is there any great impact if i tuned the kernel to "much" ? i mean the new kernel will influence the performance of the server.
3. Is there any other info or resource(s) that i can explore in order to know more on this area ?

Thank You

[Michael Tully]

Wow that many instances ....

Yes you could safely change those parameters to those settings without causing too much grief to your system depending on what else uses that parameter (maxusers) except I would not go over 4096 for maxusers.

I would also look at the shared memory pool (shmmax) and definitely look at dbc_max_pct. This setting should normally be between 300-500Mb no more. How much RAM do you have ?
Tuning the kernel using higher and higher will gradually get you to a point where you will slow a system down as it will use more resources RAM etc. Also how much swap space is configured?

[Dave Olker]

Hi Alianto,

1. Yes, you can tune nflocks to 1024 or 2048. I typically recommend nflocks be set to 2048 in environments where lots of file locking occurs, like with NFS-based applications.

I would avoid using maxusers to tune your system. There really isn't a kernel parameter called maxusers. Maxusers is a kernel formula that in turn sizes many kernel variables. I usually recommend to size the variables affected by maxusers individually, as this gives you much more control. Also, maxusers has been obsoleted in later versions of HP-UX. I just checked my 11.23 box and maxusers doesn't exist.

2. Absolutely! I've seen many cases where customers tune their systems right into the ground. If you are not sure what a kernel variable does, I wouldn't mess with it until you find out what it does or are specifically told to change it by HP or by a representative of your software appliation vendor (in your case, Oracle).

I've seen numerous systems where kernel parameters were sized way out of proportion and the system ran terribly. Sizing parameters too high wastes kernel resources and system memory.

An example: one customer set their ninode parameter at 400,000. They didn't realize that this parameter is used to size numerous different variables, and so sizing ninode in this way gave them:

an HFS inode cache of 400,000
a CacheFS inode cache of 400,000
a Directory Name Lookup Cache of 400,000
a VxFS inode table of 400,000
an NFS rnode cache of 400,000

This customer wasn't even using HFS or CacheFS, and this system was an NFS server so they didn't need an rnode cache of 400,000. Their kernel was consuming so much memory that their applications had no memory to run.

3. There is tons of information available on kernel parameters. http://docs.hp.com has many white papers on kernel parameters. The DSPP site: http://hp.com/dspp has many kernel parameter papers. The ITRC has numerous pointers to kernel parameters. Even the SAM application has a help facility in the kernel tuning portion that explains what most of the parameters mean. Additionally, any 11.23 system has a tool called kctune that can display the meaning of every kernel parameter and show you the acceptable range of the parameter.

Good luck,

Dave

[Zeev Schultz]

Here you go

http://docs.hp.com/hpux/onlinedocs/939/KCParms/KCparams.OverviewAll.html

[Rita C Workman]

No you can increase your nflocks to 2048.

You don't mention how much RAM you have. But 11 instances...wow...!

That many instances of Oracle, I might tune the basics as the others have mentioned. Check to ensure my semaphores account for the what ALL the Oracle instances together would need.

And then I would be looking at basics - where things are on disk...and how the DBA set up things like logs.
Remember, if your binaries and logs are on the same disk...ouch. If you've got 11 instances using the same binaries - that might be an issue (not sure, never ran 11 instances of Oracle like this).
I've seen a DBA put 3 copies of the Oracle control logs...all on the same disk (dah...), not to mention the archives were there too...(..hello..). Bad habits like this will clobber performance.

Another thing you might check is how things are laid out...can you take advantage of some of the different mounting options for file systems and possibly by-pass the buffer cache to help performance..(..love that mincache=direct and convosync=direct).

Often times going back to your foundation and cleaning up where things were put, and how they get mounted does so much to improve performance.

Here's another thought...are you running anything else on this box ? Running any java based apps on this box ? Java is a real hog and requires some checking...there are utilities to help with doing this.

Just a couple thoughts,
Rita

[Alianto]

Hi All,

Thanks for the feedback!

Currently the server has 2.5 GB of memory and 5 GB of swap disk.
All instances are configured less than 100MB and they are put in different logical volume or volume group.

Cheers,

Anto