What is disk striping...

[Eric Antunes]

I'm reading a note about Oracle DB_FILE_SIMULTANEOUS_WRITES that talks about increasing this parameter from it's default value of 4 if there is disk striping...

How can I know if I'm using disk striping?

Thanks,

Eric Antunes

[Bharat Katkar · Accepted solution]

Eric,
You can see it at Logical Volume level as below:

# lvdisplay /dev/vgxx/lvolx | grep Stripe

If it returns "0" then it is not striped otherwise it is striped volume.
Hope that helps.
Regards,

[Eric Antunes]

I got:

hp440# lvdisplay /dev/vg01/vol1|grep Stripe
Stripes 0
Stripe Size (Kbytes) 0

So, I've no disk striping. But I still don't know what is it?

Many Thanks!

[Bharat Katkar]

Eric,
Striping is called as RAID level 0.

What it does is it distributes the data on different disks depending on the stripe size. Suppose your stripe size is 4MB and your are stripping it accross two disk then after each 4MB write it will move to another disk and then back to earlier one and so on. This increases the I/O performances as far as writing to disks is concerned.

Regards,

[Fred Ruffet]

Strippin is a kind of software RAID5 without parity. If you have n disks and strip a LV on all of them, writing n logical extents will write 1 logical extent to each disk.

It doesn't bring security (no parity) but a better disk access (all disks accessed, instead of one). 1 write to N disks is considered better than N writes to 1 disk...

Regards,

Fred

[Bharat Katkar]

Eric,

Nice pictorial explanation on the link below:

http://www.pcguide.com/ref/hdd/perf/raid/levels/singleLevel0-c.html

Hope that helps.
Regards,

[Eric Antunes]

Many thanks guys!

I'm trying to assing points to your answers but had no sucess until now...

[Hein van den Heuvel]

All the answers so far are fine, but they focus on a detail of the question, not on the intend.

The intend of the 'if there is disk striping' is: if there is more than a single disk behind your db files.
That word 'striping' here is used in a sweeping simplyfication.


So LSM Striping as answerred so far applies.
But LSM extent based striping is fine also.
And controller based striping,
And Software (or Hardware) RAID-5
And a JBOD behind a write-back-caching controller fits the bill.
Only a straight JBOD, plain 'ol disk on a plain 'ol controller does not apply.

But then, this parameter is instance wide. So a few DB files may live on a JBOD (because thy have low IO needs?), while others live on some storage that can handle more than a few IOs in the pipeline.
IMHO you always want to increase the DB_FILE_SIMULTANEOUS_WRITES parameter.

Hope this helps,
Hein.

[Eric Antunes]

Hein,

I didn't understood what is LSM and JBOD but I'm using RAID 1 (hardware mirroring): do you think I need to increase DB_FILE_SIMULTANEOUS_WRITES from the default value (4)?

Thanks,

Eric Antunes

[Hein van den Heuvel]

Ooops, sorry. I wrote LSM (Tru64 Unix) where I intended LVM (HPUX).

The JBOD refers to simple disks (Just a Bunch Of Disks), those that you can hold in your hand. Not a 'virtual disk' such as so many storage solutions provide (EMC, XP, EVA).

For the purpose of this discussion I would qualify RAID-1 as a single, unstriped disk, when the RAID-1 is build on top of two simple disks. (PV's in LVM). But if the RAID-1 members are really chunks of stripesets, then consider it striped again.

If all your DB mountpoints/files are on simple RAID-1, then just leave this parameters.

Also, it would be nice to know (Oracle Statspack) whether there is even a problem.
Maybe writer are doing just fine (low buffer busy waits) and speeding them up may get in the way of read IOs.

"what made you consider to turn this knob"?
Just because it is there? Unless you can clearly identify a write issue (slow checkpoints, buffer busy waits,...) you may just want to leave it alone.

Cheers,
Hein.

[Eric Antunes]

Ok,

Because, after an experience I did when participating in another thread with the script in attachment (fru2.sql), I began to think that I may have an I/O contention problem. The script takes 32 minutes to execute on my environment...!!

I have tried to turn to async mode but I cannot mount the logical volumes with convosync=direct and mincache=direct options: it tells me that those are invalid options... So I'm just using rw,suid,delaylog,nodatainlog. I read somewhere that those options are just for filesystems with primarily random access... How can I know what are my filesystem caracteristics?? I'm using JFS 3.3...

How can I now if Oracle is using asyncronous writes (I've done all that is said to do in Metalink Note 139272.1)??

I'm going to see this statspack thing.

Thanks,

Eric

[Eric Antunes]

Meanwhile, I'll attach the output of the following v$system_event script in the next reply:

select *
from v$system_event
order by total_timeouts desc

[Patrick Wallek]

To use the mount options you mention (convosync and mincache=direct), you MUST have Online JFS (Advanced JFS) installed.

http://software.hp.com/portal/swdepot/displayProductInfo.do?productNumber=B3929CA

[Eric Antunes]

Hi Patrick,

I have JFS:

hp440# swlist | grep -i jfs
B3929CA B.03.03 JFS 3.3 Filesystem For 11.00
B5118CA B.03.03 JFS 3.3 Filesystem For 11.00

I don't know if it's Advanced JFS...

[Hein van den Heuvel]

Yeah, that script would create a good few writes. Many of them, I suspect, to the UNDO (RBS) rollback segment tablespace. The REDO logs, would also be busy, and lastly the data space itself.

Is that query at all representative of what the system has to do day-to-day?

You can avoid most/all UNDO activity by commiting every so often (1000?, 10000? rows).

Indeed do check out 'statspack'.
You will like it.
Then change your script to do the 'execute snap' instead of 'select time from dual'.

Also, check out the SQLplus statements "SET TIMING ON". You'll like that. Promiss!

Hein.

[Eric Antunes]

Well,

On SAM, I get:

On Source? Contents of bundle "B3929CA":
Yes JFS,r=3.3,a=HP-UX_B.11.00_32/64,v=HP
No OnlineJFS,r=3.3,a=HP-UX_B.11.00_32/64,v=HP

So, I assume I have not OnlineJFS...

Maybe I have a fragmentation issue...?

[Eric Antunes]

Guys,

I've just saw that I'm not using "Use Contiguous Allocation" for the logical volume: is it better to use that option?

Many thanks,

Eric

[Fred Ruffet]

Eric,

You don't have OnlineJFS. It should be displayed clearly in "swlist -l bundle" output. So you can't put convosync and mincache options in your mount order.

I feel more concerned as long as I gave you the fru2.sql script :)
Note I gave you that script cause you told us that you where having poor perfs on a huge delete action. I can't find the thread anymore.

Regards,

Fred

[Eric Antunes]

That's right Fred, you crashed my ego down! I used to think that I was an excelent DBA before knowing about ITRC. Now, I consider myself just a not bad DBA... :(

[Eric Antunes]

Can I install statspack on Oracle 8.0.5?

While I don't install statspack, I ran the utlbstat/utlestat and here is the output in attachment. Is there anything bad?