Configuration tips for Performance in Oracle Application Server

[Kenan Erdey]

Hi,

We have a rp3440 machine running 11.11. On this machine Oracle Application server is running.(java 1.4.2.09) We have a problem about cpu and memory usage. this is sar -5 5 output:

18:01:48 %usr %sys %wio %idle
18:01:53 59 36 0 5
18:01:58 54 28 0 18
18:02:03 62 35 0 3
18:02:08 54 22 0 23
18:02:13 53 27 0 19
Average 57 30 0 14


%sys seems to be high.when i look at java process' system calls, the "poll" call consumes much of the time. is this the problem and using async io etc. solves the problem ?

and as a solution to the memory problem. i am thinking reducing the vx_inode so that kernel uses less memory. dbc_max_pct is already %5.

lastly, propositions about configurations/best practices of application servers are welcome.

Kenan.

[Rita C Workman]

Well, some polling (depending on the appl) is normal. But if it is doing extreme polling, then you might be able to get around it. We ran into this awhile back - oracle db on HPUX, but appl sitting on Windows server. Result was extreme polling filling impacting CPU. There is a fix of sorts for this type of condition.
But a couple things would be helpful first.

1. What is your phys memory? And how about some info on your current parm settings?

2. Run tusc against an oracle process and see if you can grab enough to determine if the "polling" is really! excessive.

Rgrds,
Rita

[Kenan Erdey]

Hi,

i checked the server again. and see sys time is more than usr now. server has 8gb of physical memory. i run tusc, redirected the output to a file made a grep and see 5900 lines of totally 6900 lines about poll. i am sending kernel parameteres in attachted file.

[Bill Hassell · Accepted solution]

Poll is a programming construct which means this is what the programmer wrote. Whether this is a poor design depends on the programmers skill. In the poll man page, there is an example of poll in a program that waits 3.5 seconds before calling poll again. Perhaps the programmer wnated 'instant' response and only waited for 10 milliseconds. Ask your Java programmer. If this is a commercial application, contact the vendor and ask that the inefficient code be rewritten.

Note that while this could cause very high system overhead (as you've seen), if things appear to be working at the expected performance levels, it may not be worth fixing. Memory usage is not a problem until you run out and vmstat reports that page outs (po column) are 2 or 3 digits all the time.