Better performance

[PSS SYS ADMIN]

Hi everyone,
I would enhance application performance in a HPUX 11 environment.
I have an L3000 with 2 CPU, 2Gb ram and EMC Symmetrix disks.
Do you think I can enhance system performance modifying kernel parameters?
In attachment you can find a dump of the kernel parameters.

Regards...
PSS

[Stefan Farrelly]

The only kernel parameter which will make a big difference is; dbc_max_pct I suggest you reduce it from 25 (25% of 2Gb = 500Mb) to 15% (300Mb). If youre short of memory you could reduce it further to 10%. Not many other kernel parameters will make a big difference.

[PSS SYS ADMIN]

The system uses hardly the EMC Symmetrix disks.
There's a way to override disk bottleneck?

Regards...
PSS

[Sandip Ghosh]

You have not given any performance data. Are you facing any performance bottleneck?
Apparently looking after your Kernel Parameter it seems that you need to reduce the dbc_max_pct to 15% and watch the performance of the buffer by sar -b . %rccache should not go below 90%.
Anotherone is shmmax. If possible increase the Max Shared memory segment to 512 MB. Otherwise everything looks fine. If you face any bottleneck then post that data over here.

Sandip

[Ted Ellis_2]

hard to make really intelligent recommendations without seeing some operation and performance info, but out of the gate I see a couple of things that you may want to consider, though they are more for resource allocation that straight up performance improvement:

1. reduce the max_dbc_pct from 25 to 10... more than enough and that memory can be better used.
2. consider increasing the semmns... we set ours closer to 1000 for a modest database
3. Maybe increase the shmmax... then you can allocate a larger chunk of shared memory to a database SGA... that could have a significant performance improvement.

If you have some budget $, you can get refurbished meory sets (1 GB) from a number of certified HP resellers for about $2000 a GB.

feel free to send some performance specs to get more feedback

Ted

[James R. Ferguson]

Hi:

Before you ask to improve performance, you need to decide *where* improvement is needed.

Are you suffering from high CPU utilization and/or deep run queue lengths; high rates of memory pagin; high I/O queue depth and high device utilization?

Most of the kernel parameters are fences or limiting values. They prevent one user or one process from monopolizing a resource (e.g. a maximum number of open files).

If you are not bumping into these fences, why worry(?)

As for the infamous buffer cache ('dbc_max_pct' and its companion parameters), the use of this is best discussed based on *what* constitutes your application.

If your application is, for example, Oracle-based, then a large buffer cache is "double-buffering". That is, Oracle manages its own buffers, so why let the Kernel do it too(?). Conversely, an application that offered no real file buffer management might benefit greatly by a large number of file buffers lying around in memory.

Regards!

...JRF...

[PSS SYS ADMIN]

The application is the datawarehouse SAS.

Regards...
PSS

[Shannon Petry]

What is the bottleneck? or is there a blttleneck?
If the EMC is not being used then I doubt it has anything to do with anything, and will really be unaffected by kernel paramaters anyway.
But in your last 2 posts you had a CPU100% problem, and thought kernel tuning would fix it. Then someone recommended some bad params and you blew up your system. with them.
Kernel tuning is NOT a magic bullet and should not be mistaken for one. It will have some impact on performance but by the sounds of the severity you are having tuning wont do squat.

Really bottlenecks are self evident if you know how to monitor a system.
I.E.
CPU at 100% and application slow, disk I/O is low and memory utilization low then buy more CPU's.
etc...
etc...
etc...

No offense, but you sound like your a bit new to the UNIX market. I'll recommend that you buy a book on performance tuning for HP-UX, or take some HP Classes. While we at the forums may be able to pull enough information from you to help with this problem, what about the next and next?? Believe me as sites grow so do the problems and complications.

A good tool to let you see if there really are bottlecks and isolate them is Glance.
If you install from media you will be able to use it for 30 days free. After that you need to purchase it.

Regards,
Shannon

[Sridhar Bhaskarla]

Hi,

L-class with 2 GB and 2 CPU - seem to be like minimal configuration.

Load glance+ trail version quickly and observe where you are seeing the bottlenecks. Or look at the man page of "sar" and run few iterations of -u, -d, -q options and try to understand their output.

Depending on your application, buffer cache may sometimes help you. But it is really dependent on how your application works. You can keep it to around 300MB if you need to use buffer cache.

Enable delaylog and nodatainlog mount options on your file systems. It will improve the IO performance a bit. If you want to bypass the buffer cache as your application is already having it's own buffer, then you can use OnlineJFS's mincache=direct and convosync=direct options.

There is a performance paper published in the itrc. Search for it. You will get a lot of information and flow chart for identifying bottlenecks.


-Sri

[Bill Hassell]

It's really important to understand that trying to tweak the OS to fix an application problem. If you don't run the app, then you won't have any bottlenecks. As simple shell script can saturate any number of disks. The script wastes disk bandwidth just like poorly written SQL procedures.

As mentioned, you can adjust the buffer cache and perhaps see a 10-20% change, assuming you have a repeatable procedure to measure the difference. What will make more of a difference are the mount options but you'll need the Advanced JFS package to control mincache and convosync.

The real performance benefit will come by changing the application to stop wasting disk I/O and perform internal buffering, use indexes rather than serial reads and so on.

[PSS SYS ADMIN]

In attachment the bmp print screen of the Glance plus performance monitoring.

Regards...
PSS

[Adrian Horne]

If your bottleneck is disk and you are using SAS, is the problem the disks where your data is stored or the disks you are using for work areas? SAS can be very I/O intensive on the work disks - this is dependent on how well SAS is programmed.

Do you spread the load of work areas over multiple disks and channels? I would tend to look at tweaking usage of disks before doing anything with the kernel.

[PSS SYS ADMIN]

The problem is in the disks belonging to the SAS data vg.

Regards...
PSS

[harry d brown jr]

PSS,

that attachment is a beautiful black screen.

live free or die
harry

[harry d brown jr]

PSS,

You need to provide a little more info, otherwise, as the number of posts demonstrates, this effort is futile.

What disks are having supposed IO issues?

What is the application?

How are the disks connected to the L-class server?

How many IO paths are there to the EMC Symmetrix? Is this a SAN environment?

How many disks in the Volume Group?

How many users? Processes? Transaction Rate?

Can you purchase more memory?


live free or die
harry

[PSS SYS ADMIN]

Glance+ print screen repost

Regards...
PSS

[PSS SYS ADMIN]

Glance+ print screen repost

Regards...
PSS