kernel dynamic memory issue

[shabbir zakir]

hello,
I have an L class server. we have 4.0 gb of memory. however the free memory currently is on 34 mb. On running the kmeminfo, dynamic memory usage of kernel is 1.3 g which is high out of which malloc is using 402.8 mb and kalloc is using 899 mb which is high. I have installed the patches to fix the issue as per HPs support. Is there anybody out there who had experienced the same problem. I know on the other systems the dynamic usage of kernel memory should be only about 300 mb.
Please help
thank you
shabbir

[Steven E. Protter]

1) Having a suppor tcontract is a good idea.

2) attach output of a kmtune command.

3) Maybe try some data collection over time to get and analyze performance data. See the script I attached.

P

[Sridhar Bhaskarla]

Hi Shabbir,

In the Physical memory usage summary of kmeminfo, how much are you finding for Bufcache?.

I believe what you heard was about the dynamic buffer cache which is to be set to around 300MB.


-Sri

[Gary Yu]

Hi,

What's your value of kernel parmerter "dbc_max_pct"? its default value is 50, which will take 50% of your physical memory as dynamic cache, you should change it to 15.

Gary

[Wodisch]

Hi shabbir,

with your RAM size even 10% would be fine for the UNIX buffer cache.
How about all those tables? ninode, nproc, nfile, nkthread, etc.
I can only guess they are huge...
Maybe you don't need those set that high?

FWIW,
Wodisch

[shabbir zakir]

hello,
Here the output of kmtune,kmeminfo and sysdef
the HP support has turned it over to the labs.
I am going to run simple performance (sar,vmstat) any ideas how to get hold of vmtrace/tusc/q4.

[shabbir zakir]

Hello,
The first file was kmeminfo(0203)was 24 hrs after the reboot.while 0131 was just after the reboot.
heres the other data

[shabbir zakir]

Hello,
heres the sysdef

[James Murtagh]

Hi Shabbir,

The output you provided doesn't really tell too much. If you are talking about a kernel memory leak you need to provide "kmeminfo" without any options. You mentioned HP provided patches - did that confine/fix these leaks? If the memory usage is from applicaction usage there is little HP can do.

Regards,

James.

[Dietmar Konermann]

Pls. post the default output of kmeminfo also. (w/o -user)

[shabbir zakir]

Heres the default kmeminfo

[Dietmar Konermann]

The most kernel memory is consumed by bucket[11] and the superpage pool.

PHKL_28172 may help to prevent fragemtation of the superpage pool. The bucket[11] usage may be caused by a large VxFS inode cache. You could try to set the vx_ninode kernel tunable to say 20000 as a test.

These are only guesses. For a detailed analysis you should log a call with your local response center.

Best regards...
Dietmar.

[shabbir zakir]

Hello all,
Thanks for the tip. I tuned the vx_ninode to 16200 and vx_noifree to 1 and so far the system is showing that total dynamic memory usage is 198 mb which is very good. and also the bucket 11 and kalloc are also pretty low.
By the way one question could anyone tell me how does all these information from kmeminfo correlate to different systems like vxfs etc. or if you direct me to good source of OS/kernel internals understanding.
Very nice forum and hope to learn new things every day.