Striping and Performance

[Popy]

Hi Gurus

I have a filesystem of 125 GB (approx) created with 11 PVs.This is where my Oracle lives and of course very intense read-write happens.Somehow, we missed striping while creating LVs..my current stripes says 0 .

Now, my question is, if I do a re-build of my LV with striping enabled (say 10 stripes )what improvement I can expect ? And, what is the impact of stripe size while we do set the striping ON.
Here is my current LV details,

LV Name /dev/vg02/lvol2
VG Name /dev/vg02
LV Permission read/write
LV Status available/syncd
Mirror copies 0
Consistency Recovery MWC
Schedule parallel
LV Size (Mbytes) 140800
Current LE 35200
Allocated PE 35200
Stripes 0
Stripe Size (Kbytes) 0
Bad block on
Allocation strict
IO Timeout (Seconds) default

Many Thanks,
Raneesh Vijayan

[Luk Vandenbussche]

Hi,

It is difficult to tell what the improvement will be. But is will be beter. That is your disk array?

One thing is sure, if you want to enable striping on a non striped lv, you loose all your data in the LV.
So take the following actions
1) Backup data
2) create a new striped lv
3) restore data

[baiju_3]

Hi Vijayan,

The existing LV can not be converted as such , you need to reacreate it using lvcreate command. So you have to go for a backup of your data and restore it after establishing the striping .

lvcreate -L 500 -i 10 -I 64 /dev/vg03

The above command will create a 500MB LV with 10 disks and 64KB as stripe size.

Striping will definetly improve the I/O performance .In our environment we keep the stripe size of 2048 (2MB) for a 10 disks.
thx,
bl.

[Geoff Wild]

We stripe - but not at the lvol level - we let the array take care of that - striping should increase performance - we stripe across the whole array - reads/writes are faster as you have more heads accessing the data.

If you have 11 PV's - then stripe across 11....

Rgds...Geoff

[Rita C Workman · Accepted solution]

I agree that striping will give you some improvement...but...
I too inherited Oracle instances of various sizes and none were on striped disk.
You can improve the performace by doing some simple things. Let me rephrase that, simple but involves work even on non-striped disk. So if you don't want to redo as striped..maybe..

Look where files are. Can you move some lvols to different disks on seperate spindles that will help improve performance.
I mapped out every lvol to disk and noted what file types were there.

What about your mount options. Are you wasting time double buffering? If you set your mount options right on your file system you don't double the work that Oracle is already doing. First ensure that what is under each mountpoint is what it should be, then set your mount options correctly.

Now under each mountpoint...what is being hit the most. Are there other dbf files (or whatever) that maybe could be moved somewhere else. Like I said...map your lvol/fs out and you may be surprised who put what - where.

Give your oracle logfiles enough space to spread out. I prefer to keep the size of the logfiles down. Better more small then less large.
Check where your oracle logs are writing..and move them around. I like logfiles created so they hit different disks. Log 1,3 hit disk-a but Log 2,4 hits disk-b. Small improvement here, but we sync to remote site..so in that case it helps. This may not apply for you.

Work with your DBA's...check that init.ora and see where they put things. Check where folks are putting files; where the dbf files are. And move what needs to be moved. Set down the rules of where things can and can not be put...period..
pvmove became my best friend when I had to try clean up disk performance issues here.

Not sure what disk hardware you have. Here they went with EMC. They have a real nifty software called Powerpath...let's you double your I/O cause it will use both paths to access the disk. Huge performance improvement with this.

Just a thought,
Rgrds,
Rita

[Popy]

Thank you all for your excellent advises,

Here is the answer for your queries,

SAN Box- XP 512
Mount option - delaylog,nodatainlog,largefiles 0 0

Cheer's

Raneesh Vijayan

[Peyman Javaheri]

Hi Raneesh,

That seems to be a very expensive JBOD. Why not stripe right in the XP array. It should provide the best performance for you compared to hostbased RAID.

regards,
peyman;

[Ted Buis]

Another point to consider is increasing the queue depth. HP-UX will by default send out 8 "SCSI" command (yes even over FC it speaks a SCSI command set). The XP512 can handle many more than this on a CHIP port. If you don't have to many servers or too many LUNs per port, you can increase the queue depth so the array can have more commands to be processing in parallel. Read the man page on scsictl and look at the "-m queue_depth=XXX" option. This is done live and won't survive a umount or reboot, so you need to build up some scripts to do this for each LUN. Don't forget to do the analysis to see that you don't exceed the queue depth that the XP512 can handle. I don't remember it exactly but it is likely to be something like 1024, but check the documentation to find out what it really is.