buffer-swap-memory

[Domingo Muñoz]

I have got a server model L2000-44 with 3GB of memory. I have opened glance and I can see the memory to 100% (with 20% buffer to used more or less), 30% user swap and 35% Reserved swap.
I have dinamic buffer between 5% to 30%.


I don??t know if this is correct because I thought the memory reached 100% then the buffer decreased to 5%.
what is the reseved swap used for?

[Steve Steel · Accepted solution]

Hi


This gives a good explanation.

http://bizforums1-it1.mayfield.hp.com/cm/QuestionAnswer/1,,0x6752a22831ebd5118ff40090279cd0f9,00.html

steve steel

[Corthouts Carlo]

On your system you can find a document in the /usr/shar/doc directory named mem_mgt.txt

It is the memory management white paper and explains everything.

This document can also be found on http://www.docs.hp.com.

For as far as I understood the usage of reserve is that the system reserves swapspace and memory when spawning processes. When there is not enough memory or swap then no process will be spawn and you get the error message can not fork.

[Hisham Ahammed_1]

hi
in my knowledge dbc_max_pct=30% means, the system will be using max of 30%of u'r physical memory for the buffercache. in your case it is not using 30%, says that u may not be having much filesystem related processes as buffercache is coming in the picture, and if u don't have more filesystems and using more raw partition this type of scenario can exist.
use this if helpful to u

rgds
hba

[Paula J Frazer-Campbell]

Hi

Seeing 100% use of memory is not a bad thing, as memory is the faster medium on your system.

Calls that cannot be dealt with by memory, either processor cache of system memory use disk a much slower mechanical medium.

If you search for settings for dbc_max_pct you will find that an answer of ???It depends???.

It depends on what your server is being used for.
A figure of about 400Mbyte is felt to be a good starting point.

Your setting of 30% gives 900 Mbyte a little too much, a setting of 15% would be better.

Sar ???b 1 100

05:00:00 bread/s lread/s %rcache bwrit/s lwrit/s %wcache pread/s pwrit/s
05:01:00 32 9561 100 2 168 99 0 0
05:02:00 6 12259 100 2 144 99 0 0
05:03:00 1 2614 100 1 274 100 0 0
05:04:00 6 14963 100 2 152 99 0 0
05:05:00 4 17052 100 1 147 99 0 0
05:06:00 1 12180 100 1 208 99 0 0
05:07:00 0 8444 100 2 176 99 0 0
05:08:00 0 14121 100 2 136 98 0 0
05:09:00 0 686 100 0 31 99 0 0
05:10:00 0 2998 100 3 61 95 0 0
05:11:00 0 591 100 3 25 86 0 0
05:12:00 0 934 100 3 65 95 0 0
05:13:00 0 674 100 1 25 98 0 0
05:14:00 0 1213 100 0 64 100 0 0
05:15:00 0 698 100 2 27 94 0 0
05:16:00 0 786 100 3 59 95 0 0
05:17:00 0 743 100 2 28 94 0 0
05:18:00 0 850 100 2 60 96 0 0
05:19:00 0 694 100 1 23 95 0 0
05:20:00 0 797 100 2 62 97 0 0

With dynamic buffer cache the %rcache is the column to watch as 100 = 100% success in finding what it wants in memory.
The aim is to get the max setting so that almost 100% is achieved but without wasting system memory.

HTH

Paula

[Domingo Muñoz]

Only I need to know if my server is 100% of memory then the buffer cache should have un 5%. Is correct?

I have dinamic buffer cache between 30%-5%.

[Bill Hassell]

No, the system will not reduce the buffer cache just because memory has reached 100%. Instead, the buffer cache is reduced when processes need more RAM. This can happen when more processes are started or existing processes ask fore more RAM.

Now 11.0 improved the algorithm to push the buffer cache down more quickly than 10.20, but 11i is vastly superior in this respect. Note also that idle programs (ones that are waiting on a keyboard input) are lower priority that the buffer cache (which is always busy) and these will be deactivated and paged out first, before the buffer cache is reduced.

[Ruediger Noack]

Hi Bill,

a question to a hp pro. You wrote:
Instead, the buffer cache is reduced when processes need more RAM. This can happen when more processes are started or existing processes ask fore more RAM.

This is the only logical explanation for me. But on the other hand here in the forum I can read a lot of suggestions to reducing of dbc_max_pct because of reserve more memory for processes. This would dramatically increase the system performance. Is it possible they are all wrong? I don't believe it. Or is this a historical experiance from older hpux versions.

Thanks for your (and others) explanation.
Ruediger

[Bill Hassell]

The dbc default value is outrageously large and in 10.20, it takes a lot of prcesses asking for more RAM before the dbc backs down, thus the recommendation to keep the dbc-max% to about 10% of RAM or less than 300-800 megs, whichever is lower.

At 11.0, the algorithm to free dbc areas and reduce the size due to process pressure is improved but the full benefits won't be seen until you upgrade to 11.11 (11i). As mentioned in other posts about performance, memory usage is not a good metric at all. Instead, watch the page-out rates in vmstat or Glance. Single digits are OK, 2 digits up to 20-50 are probably OK if it is not continuous, and page-outs in the hundreds means there is way too little RAM. A 10.20 or 11.0 system with only 128 megs of RAM cannot even run SAM without significant paging (takes 2-3 minutes for SAM to start in character mode).

[Ruediger Noack]

Thanks Bill,

your answer is very helpful for me.
I hope Domingo will assign many points for that.

Ruediger