subject: arraydsp -r says I need "another disk"

[Brad Marks]

I'm running HP-UX 10.20 on a K-420. As the 'subject' of this post indicates, I may need more disk. I have an H/P AutoRAID 12H with four 18.2 disks. I have three vg's: 1 dedicated to system stuff; 1 for programs; and 1 for data. 'bdf' says that I have 12gb free on the vg dedicated to data, I'm still told I need "another disk".

Do I need another disk or do I need to create more than one vg for my data.
Glance is also giving yellow and red disk warning messages.

Please find attached the output of 'arraydsp -a'.

Further, 'arraydsp -m' for a 24 hour period during a weekday, shows SCSI Q metric to be OutOfSpec for each and every of the 24 hourly report periods! The range in value is from 73.000 to 121.00. Is it a disk or maybe a SCSI bottleneck?

Please excuse my posting this again, but I'm hopeful that I might get more helpful responses.

Any help would be greatly appreciated.
Thanx again!

[Paula J Frazer-Campbell]

Hi

Is the glance error disk busy errors?

What does sar -d tell you on the %busy column?

What are the glance warning levels set to?

Do you have online JFS installed if so perhaps a defrag of the disk will help.

HTH

Paula

[Brad Marks]

Paula,
Thanx for the input.
-Unfortunately, I don't have onlie jfs.
-Glance Disk I/O rate threshold is set to all. What does this mean?
-sar -d says that the average %busy is 21.37%
but has moments when it is 100%.
-A few hours ago I cleared the alarm history in glance but do believe that 'busy disk' was the yellow and red warning message.

Thanks for your time and consideration.
Brad

[Brad Marks]

Here is glance's messge text:
Time Level Message
13:00:09 YELLOW Disk Bottleneck Probability=73.29%

[Paula J Frazer-Campbell]

Hi Brad

From sar and glance your disk is busy - my first try to improve it would be a defragment it, as on an active database files slowly do get defragmented and will slow the access time to them as well as increase disk activity.

Without online JFS it is not a task that can be done on live data.

First stop all access to the data disk then get a good backup of your data to another local disk/nfs disk if you have the space but if not then to tape.

Do a rm on your data (you have to be 100% happy with your backup) and then restore your data to the disk.

Once completed onCE again monitor the disk activity (glance and sar)- any improvement will depent on how bad the defragmentation was.

If your data in from a Universe database then get your DBA to do an ANALYZE.FILE on the big busy files and then a resize. DO this before the defrag.

HTH

Paula

BTW - If this data is business critical I would fit another disk in and mirror the data for safty.

[Brad Marks]

Will additional disk drive(s) not assigned to a LUN make any differnece at all?

[Paula J Frazer-Campbell]

Hi
I am not that "Au Fait" with auto raid but would imagine no - thought I may be wrong.


Paula

[Bill Hassell]

You don't need another disk. You need the performance of your system analyzed first. No matter how many disks you have, nothing will help if they are all on the same channel and the channel is saturated.

Performance basics:

There are only two resource limitations: CPU and I/O. That's it. So if the CPU is not saturated, then it's time to look at where all the disk I/O is going, and most important: is all that disk I/O necessary?

This is not a simple task as it requires a lot of data. Classic Unix tools like sar and top will be very cumbersome, requiring hours of collecting and interpretation. Since you have Glance, the job is a lot easier.

Start by looking at RAM. Modern databases need massive amounts of RAM to prevent overloading disks. For 10.20, make sure you have 4 Gb and for 11.0, bump this to 8 to 16 Gb. Then have your DBA change the database to use as much RAM as possible. In Oracle, this is the SGA.

Check Glance to see how big the buffer cache is and what the percentage of Logical reads might be. The percentage should be in the 93-95% range over several minutes. A larger buffer cache can help improve the logical read rate.

Then determine which filesystem has the most activity and what files are in use. These could be moved to another physical disk. However, a lot of this can be configured in the database.

If your array haas dual controllers, make sure the volumes are inteleaved between the two channels so that two channels can be used at the same time.

[Dan Rosen · Accepted solution]

I don't want to disagree with an HP expert. Every thing he said is true. You must analyze to see what is causing your bottleneck.

However, once you do (as we have), and depending on what you do with the file systems on the server, you may find significant gains by adding drives.

Here's why in a nutshell: depending on options, your autoraid is probably reserving a WHOLE disk for hot-swap/fast recovery. With only four disks, that is a lot of overhead. Now when a request comes into your Autoraid for data, there are only three disks that respond and begin pumping data. With a fully loaded Autoraid you would have 11 disks pumpng data simultaneously.

What you fix by having more disks is the contention for the physical heads on the drive. In our setup, using a mission critical database, it is a must. We left an ASP where we had the database on 10 fiber channel disk array, and there is a real difference between that and five SCSI-II disks in the Autoraid. (And our host computer is faster).

Also beware of adding different size disks in the Autoraid, it can cause lots of wasted space.

Can anyone with an Autoraid answer if a fully loaded Autoraid can overwhelm the SCSI-II connection to the server? I would have gone Ultra-SCSI III, but the D-class doesn't have any interface cards.

[Bill McNAMARA_1]

you can determine if you need another disk or not.
If your performance is just fine, you don't need one. If it is poor performance, yes, by all means add more disks. (or delete luns)!

Later,
Bill