Write Cache hit ratio

[Justo Exposito]

Hello,

I'm looking with patrol my system and I have a problem with the memory cache, Patrol tell me that is to low the % Write cache ratio.

How can I solve the problem?

I suspect that the problem appear since I change the dbc_max_pct from 50% to 25%. I have 2 Gb of RAM and I am runnig Oracle Database in a Cluster environment.

This is the explanation that patrol show about this parameter:
"This parameter displays the percentage of logical blocks written in the buffer cache. The write cache hit ratio indicates the effectiveness of system buffering. If you run I/O intensive applications and you have a large memory configuration, you may want to reconfigure your memory so that the write cache hit ratio is around 80 percent."

And our system is under 20%.

Thanks in Advance,

Justo.

PD:Attached is my /usr/sbin/kmtune -l -S /stand/system output.

[Michael Tully]

Hi Justo,

What is the output of glance as
opposed to the stats produced by
patrol? For example if you run
swapinfo -t it can give different
stats to what glance tells. If
you don't have glance installed
there is always the trial version
that you can install. Also what
type of disk hardware are you using?
This may also distort what could
be the problem.

-Michael

[Justo Exposito]

Hi Michael,

This is my swapinfo -t output:
HP4 #> swapinfo -t
Kb Kb Kb PCT START/ Kb
TYPE AVAIL USED FREE USED LIMIT RESERVE PRI NAME
dev 2097152 251088 1846064 12% 0 - 1 /dev/vg00/pswap
dev 2097152 252148 1845004 12% 0 - 1 /dev/vg00/page01
reserve - 1536936 -1536936
memory 1572536 655536 917000 42%
total 5766840 2695708 3071132 47% - 0 -

We have a diskarray FC60 installed, I have glance, where can I look for information (What options I must use)?

Thanks,

Justo.

[Ian Dennison_1]

I guess the Write Cache depends on how hard the Database is working. If you are doing a lot of database output in your Oracle Application, then the %write will always be low.

Are you mounting the Database File Systems with the 'direct' option (only available on OnlineJFS), which bypasses the Buffer Cache?

How about distribution of data across different Disks and Channels? Seeing high-disk usage (100%) on 'sar -d'?

As long as your overall %wio is not high, and your Read Cache is fine, and the system is not working hard to manage the Buffer Cache (I forget the name of the daemon that manages it), then you should be fine.

I quote to my DBAs that allocating more than 400MB of Buffer Cache incurs an overhead, and that the memory freed up by reducing this Cache would be better used for the Application (Oracle).

Share and Enjoy! Ian

[Michael Tully]

In glance you can see the buffer,
memory and swap information from the
F3 key. Ian gives some pointers in
regards to how the database works.
Depending on what hardware you are
using will depend on how accurate
then output of 'sar' is. General
consensus is that anymore than 10-15%
of memory or more than 300Mb of
buffer cache is too much.

-Michael

[Justo Exposito]

Hi Again,

We are loading data into the database at this time, then we are writing in disks. Then this is a normal behavior?

This is produced while the CPU rate is near 100% and we run the process that load the data into the database.

Ian I'm interesting in the direct mount option in Oracle, can you tell me more about this?

Thanks Michael & Ian,

Justo.

[Ian Dennison_1 · Accepted solution]

Justo,

The basic premise of mount direct is that it bypasses the Buffer Cache and talks directly to the Disks (and their cache). It reduces a step in the process of caching, but does highlight limitations in disk speed and layout. It also prevents the OS from using Buffer Cache for these file systems.

If you are performing a load, is it an initial load or a regular occurence?

Below are a few links to Forum entries on this subject

Oracle Setup

http://forums.itrc.hp.com/cm/QuestionAnswer/1,,0xf97e37f45ef7d4118fef0090279cd0f9,00.html

Has some good rules of thumb for setting up Mincache
http://forums.itrc.hp.com/cm/QuestionAnswer/1,,0x35bbc8ecad09d6118ff40090279cd0f9,00.html

OmniBack Backup

http://forums.itrc.hp.com/cm/QuestionAnswer/1,,0x1067ded2442ed5118fef0090279cd0f9,00.html

What sort of Disk Storage are you using? An AutoRaid or XP Array may benefit from some tuning to make this work.

Share and Enjoy! Ian

[Ian Dennison_1]

Justo,

OK, found out that you are using an FC60 array.

Question: Are you using RAID 0/1 or RAID 5?

The FC60 Array Documentation has the best example of performance comparisons for RAID Types I have ever encountered (pg 47 of the Users Guide)
http://www.hp.com/cposupport/manindex/hpsurestor19163_eng_man.html

Question: What's your cache settings ont he Array? (amdsp -i, then amdsp 'Array ID')?

Cheers, Ian Dennison

[Justo Exposito]

Hi,

Sorry for the delay, I must go out at this moment. Then continue in monday, it's very interesting for me.

Many thanks Ian, see you next monday.

Regards,

Justo.

[Justo Exposito]

Hi Again,

We are doing an initial loading, but our application will need regular loading of data, because this is not the central application and is a database server for a call-center that is Siebel.

Our Disk Array is an FC60 with a 1 Tb in mirror.

The main problem is the sql quality that Siebel generates, this is a pour SQL quality then we have a lot of table-scan then we can not use the direct option for our JFS file-systems.

Many thanks for your assistance,

Regards,

Justo.