performance difference for pdf creation

[Charles Li_1]

Hi,
We have 2 servers, one with Linux and one with HPUX. We deployed a j2ee application on both servers. This application queries the database and generates a pdf file. The HPUX machine has a lot more CPU and memory than the Linux server. But, the HPUX machine took 5 times as long to run the application as the Linux server. We would expect the reverse to be true, due to the memory and cpu difference.

Has anyone encounter this, or know what is causing the performance difference?

HPUX: PA-RISC, 2 CPU 875MHZ, 12 GB of ram
Linux: Intel Xeon, 2 CPU 800 MHZ, 8Gb of ram

Thanks.

[Mark Greene_1]

There could be a lot of things, I'd start with the kernel parameters, specifically what are dbc_max_pct and dbc_min_pct set to? Also, are the patches current? And are you looking at the actual timestamps on the files or when the files were made available to the client side? And if the latter, look to see if you have a duplext mismatch with the network connection for the HP.

mark

[Charles Li_1]

Mark,
This is what we have for the kernel settings:
dbc_max_pct is 50 ; min_pct is 5

Are these settings fine, or do we need to change them? What are these settings use for?

Our OS patches last done at the end of September.

I am timing the application on when I press the report button and when it comes back with the pdf report on my desktop.

Thanks for your help.

[Mark Greene_1]

You'll want to change the kernel settings to 10 and 2, respectively. There are other things with the kernel you can tweak, but these are generally the ones to start with.

I'd also suggest running pings and traceroutes from both boxes to your desktop to see if there's a disparity in the network performance and number of hops. The duplex match-up with the HP system and the network switch to which it's attached is also a key item to verify.

mark

[Charles Li_1]

We try the kernel parameter setting change it is still the same. One thing I forgot to mention is that when we up the heap size to 256MB, from the default of 44MB, the performance got better. In speeded up to 1.5 minutes from 5 minutes. But it is still slow comparing to the Linux box as 50 seconds, configured with default heap size of 44 MB.

[Hein van den Heuvel]

>> We try the kernel parameter setting change it is still the same.

Of course. dbc settings are rather unlikely to have a major effect for an application which is unlikey to do major file IO.

>> One thing I forgot to mention is that when we up the heap size to 256MB, from the default of 44MB, the performance got better. In speeded up to 1.5 minutes from 5 minutes.

Excellent. That is progress.
What I miss most in your description is an indication of what is happening during those 90 seconds on the HPUX box versus the Linux box.
- Are those systems 100 busy for 1 cpu (50% total cpu)
- Is there perhaps significant network work and is the network efficiency different from the selected client. How about ftp-ing a dozen similar files from both hosts to the client and see if that performs similarly.
- If there is significant cpu burn, then which process is burning it, and which mode?
-- User time or System time
-- Java provess or Database process.

== If it is high cpu time in the java process, then further java (GC) tuning my be your best bet.
If it is the database,

== If it is the database where the time is spend, well then is it time for basic sql analysis. Get a trace, compare execution plans, verify indexes and so on.

Cheers,
Hein.






But it is still slow comparing to the Linux box as 50 seconds, configured with default heap size of 44 MB.

[Charles Li_1]

This is what I see using top:
On the Linux server, the CPU usage jumps to 99% immediately. Whereas the CPU usage for the HPUX server went up steadily, to 99%. It took about 20seconds to it to reach 99%.

I am using the same client/PC for this test and they are using the same DB.



- If there is significant cpu burn, then which process is burning it, and which mode?
>> can you tell me how to find the mode that you refer to

-- User time or System time
>> I am using my watch, timing it from when I click the report button until the report appears on my pc

[Hein van den Heuvel]

>> On the Linux server, the CPU usage jumps to 99% immediately. Whereas the CPU usage for the HPUX server went up steadily, to 99%.
It took about 20seconds to it to reach 99%.

Wild speculation on my part, but to me this _suggests_ that on HPUX the database (Oracle?) still needs to find stuff, from the disks, before it gets into using that data. Are the databases configured compareably? Similar buffer space (SGA?)? Same indexes? Tables analyzed and with valid stats in both places?
Of course it could also be growing address space, or doign security og entries, or maybe a database trace is active on one box and not the other. So many possible differences to eliminate!

>> and they are using the same DB.

How can they be using the same DB?
Different box, different architecture no?
It may be the same software & version, and you may have attempted to load the data exactly the same, but difference all too easily sneak in.

>> can you tell me how to find the mode that you refer to

Just using TOP or VMSTAT (maybe with 1 second report interval for the duration of your test). Look at system and user time. You may even want to post vmstat output for both systems here as a .txt attachment.
It woudl also give an indication of memory management activities (if any).


-- User time or System time
>> I am using my watch, timing it from when I click the report button until the report appears on my pc

Again, SAR or VMSTAT can give a first explanation on how the server time is being used and whether there are recognizable phases.

Is this a client/server setup where the processes stay around? In that case a ps -ef before and after and calculating the cpu time increments may give a nice hint at to where to look, but normally TOP is quicker.

hth,
Hein.

[Charles Li_1]

Sorry, I might not be clear on the setup. The same j2ee application is running on separate server which uses the Oracle Application Server (j2ee container and Apache). These 2 servers, HPUX and Linux, runs the same application and connects to the same database.

[Hein van den Heuvel]

Ah! Much clearer now, but moving into an area where I have less value to add.

So this is a 3-tier setup really.
And like you said the DB is the same.
(barring perhaps a minor details in SQLnet configs and TCP detail settings as sometimes defined from the middle tier).

If you have insight/connectivity to the DB backend i would still recommend getting an insight into when, and how long, it gets invoked, but i now suspect that not to be too relevant.

Sometimes for a quick shootout I would again use vmsstat in two windows, one on the middle tier, one on the db at 1 second rate to understand the timings/loads better.

Back to Oracle apps server, and java tuning...

good luck!
Hein.