wcache low in sar

[Javier Ortiz Guajardo]

I??have a HP k570 with 3Gb of memory, hpux 11.0 64 bits...
i??ve already install glance (the try) and it??s expired.

I??m running sar -b and i attached the output.

The %rcache is fine in 100% but the %wcache falls below the 75% (some times is 22) and i??ve been reading that this is a pour percent.

the parameters for buffers are
nbuf 0
nbufpages 0
dbc_min_pct 5
dbc_max_pct 50

also the command
$ echo bufpages/D|adb -k /stand/vmunix /dev/kmem
bufpages:
bufpages: 59676

wich parameter i must modify to make %wcache working properly?

i??have a serious problem with swap too
$ swapinfo -mat
Mb Mb Mb PCT START/ Mb
TYPE AVAIL USED FREE USED LIMIT RESERVE PRI NAME
dev 1800 630 1170 35% 0 - 1 /dev/vg00/lvol2
dev 3000 626 2374 21% 0 - 1 /dev/vg00/lvol9
dev 1200 622 578 52% 0 - 1 /dev/vg00/lvol10
dev 3000 634 2366 21% 0 - 1 /dev/vg01/lvol1
dev 3000 624 2376 21% 0 - 1 /dev/vg01/lvol2
dev 3000 632 2368 21% 0 - 1 /dev/vg02/lvol1
dev 3000 628 2372 21% 0 - 1 /dev/vg02/lvol2
dev 3000 632 2368 21% 0 - 1 /dev/vg02/lvol3
reserve - 982 -982
memory 2329 404 1925 17%
total 23329 6414 16915 27% - 0 -

what should i do?.
can i tune some of the kernel parameters to gain memory resources?(wich ones)

I need very much your help. and thanks so much?

[Paula J Frazer-Campbell]

Hi
Your :-

dbc_min_pct 5
dbc_max_pct 50
Are at default settings

It would be better at approx 80 meg for min and 400 meg for Max

dbc_min_pct 2
dbc_max_pct 13

A reboot is required.


Paula

[A. Clay Stephenson]

You need to understand that "wcache% below 75% bad" is nothing more (and nothing less) than a "rule of thumb" - it always comes with an "it depends" attached. For example, if you are updating a database where lots of random i/o occurs and little sequential i/o occurs then your results would be perfectly normal.

The most revealing aspect of your data is that you have a good bit of block writes but zero logical writes. This strongly suggests that you are running raw/io; in that case, your results are perfectly normal. In raw/io (or the onlineJFS equivalents convosync=direct,mincache=direct), buffer cache is bypasswd so that no amount of buffer cache is going to improve the ratio.


Typically 10.20 and 11.0 boxes perform best with buffer caches set to around 300-400MB; 11.11 works well up tp about 1GB. I would cerainly reduce to dbc_max_pct so that it is no more than 400MB although I tend to run fixed caches by setting bufpages to a nonzero value. 80000 would be a good starting point (320MB).


The REAL questions that you should be asking are: "Is my system performing poorly? Are
the users complaining?"

[Javier Ortiz Guajardo]

Thanks so much.
Clay.

We have running only SAP in this machine, as you kwon SAP/R3 has its own parameters.

And the statistics show that the response time has been increased. And the machine is swaping.

Just for this, my boss is asking me if i have some parameters that i can reconfigure to gain some memory. So the SAP can take it.
And reduce the amount de use of swap device.

We are checking the SAP parameters too.

but i have to much trouble to measure the parameters from kernel, i??ve already check the nfile, nproc and ninode parameters and i think i have this parameters configured high. and i think we can have more parameters in the same way.

that??s why i'm asking for help to tune my machine.

Thanks in advice.

[A. Clay Stephenson]

Sorry, I read the top part of your question and I completely ignored your swap. The swap is your real problem; wcache% concerns in this case are completely trivial. You really have only a few choices: 1) reduce the number of processes; 2) reduce the shared memory usage; 3) lower buffer cache but in your case it ain't gonna help 4) Reduce ninode - it really only applies to hfs filesystems and I'll bet /stand is your only one (~1000 should be more than enough); 5) Buy more memory - that's the only real option.

There is no point in running swapmem_on=1 when swap space exceeeds real memory. You can turn it off and reduce a little overhead.

Your real goal is to reduce pageouts (swapping) to very, very near zero. You can purchase 3rd-party memory at a fraction of the cost of HP.
It's also time to purchase Glance, like MirrorUX and OnlineJFS, it's one of those tools no HP-UX box used for any serious purpose should be without. The cost of not having them (like not having enough memory) greatly exceeds the cost of the product.

Run vmstat and look at the po (pageout) column; it's the only column of any real use in the command. If that rate is not nearly 0 then you have swapping problems.