disk upgrades

[Jennifer Young_1]

I have 16 4GB drives in 2 HAAS A3312A (Jamaica) units. Data is mirrored onto another 2 units. I am getting some serious bottlenecks, the setup of the filesystem is 1 filesystem per disk. What i am wondering is if i switch to 18GB or 36GB drives in a different array, say, the hvd10, and stripe all the data across higher speed disks, will this have better i/o performance even though fewer disks are being used?

here are some additional questions i am asking myself:
should i use the same setup with the hvd10 as with the jamaica, ie, using a separate unit for mirroring? or should i replace our 4 HAAS units with 1 hvd10 populated with 36GB drives, and use half of the unit on a separate controller for mirroring? or will this decrease i/o performance too much since only 5 disks are being used, compared to 16 on my current setup?

thanks for any input, id like to hear from anyone who has been in a similar situation. btw, i always assign points.

[Curt Thompson · Accepted solution]

Hello Jennifer,

Intuitively, it would seem that if you used higher RPM disks, fewer of them per SCSI channel and striped the data across them that I/O performance should improve. However, to be absolutely sure, you will have to do the experiment. With all those things going for you, I think the performace will improve (probably since not all the filesystems get the same amount of I/O, and the ones that do get high I/O will improve dramatically with striping).

Also, if you can swing it, I would go with multiple HVD10's for redundancy purposes. Although the HVD10 is a very hot-swappable device, the midplane board (located in the center of the HVD10; where all the disks and controllers plug) requires the whole unit to be down for replacement.

Good Luck,
Curt

[Charles McCary]

Jennifer,

do the disks in any of your arrays support alternate / dual paths? If so, you should make use of these when configuring the volumge group than a particular file system will reside in, along with striping a particular lvol across multiple disks.

For example if a given disk is accessible through c0t0d0 and through c1t0d0, make sure when you add the disk to the vg that you add both paths. Also, if you have multiple disks with two paths, make sure that you alternate which disk is the "primary" path. For example if you were adding 4 disks to a given volume group:
Path 1 Path 2
disk 1: c0t0d0 c1t0d0
disk 2: c0t0d1 c1t0d1
disk 3: c0t0d2 c1t0d2
disk 4: c0t0d3 c1t0d3

you would want to add them as below:

vgextend /dev/vg_whatever /dev/dsk/c0t0d0 /dev/dsk/c1t0d0 /dev/dsk/c1t0d1 /dev/dsk/c0t0d1 /dev/dsk/c0t0d2 /dev/dsk/c1t0d2 /dev/dsk/c1t0d3 /dev/dsk/c0t0d3

Notice that I alternate which path is first as I go down the list, this assigns the primary path in the volume group.

Now, after you've done this, if you stripe an lvol over all 4 of the lvols above, you've gotten about as much performance as you're going to get out of these 4 disks (if they are dual path as explained above).

Hope this helps.

[KapilRaj]

Hi i will give you a solution ,

I take the following things as assunptions

01. You are not worried about the Filesystem capacity.
02. You are more worried about the performance.

For achieving the above out put you do not need to replace anything. Just keep it as it is.

You said "One filesystem per disk " . which effects the performance very badly.

I would suggest you the following practice.

-Take a full system back (all & all)
-Make a single vg for all the disks
-Make two PVGs (Physical volume groups, one pvg which consists all the disks in HASS 1 & 2 other consists HASS 3 &4
- lvcreate -LXXXX /dev/vgXX (Create all logical volumes like this)
- lvchange -s g -D y /dev/vgXX/lvolXX (turn on strict PVG & Distribution)

This setup will strip your Logical volumes into number of disks ,which will improve the I/O .

If you are worried about space. Go for 36 GB modules but increase the number of disks so that Diskdrive access time is not a bottle neck.
Same time try to go for seperate SCSI channels

Get in touch for further inputs.

Wish you all the best,


Kaps

[John Waller]

Jennifer,
The thing to bare in mind is that by reducing the number of disks you are moving the "bottleneck" away from the SCSI Bus to the disks themselves. I suspect your problem at the moment is due to a bottleneck on the SCSI bus rather than disks, you have not mentioned how many controllers you are running with, but 6 - 7 disks per F/W controller is really the maximum recommended for good performance.

You have not mentioned if you have to work to a tight budget, but I would recommend that if you move to the HVD10 solution, I would install 2 (1 + 1 mirror) running with 20 disks but setup in split bus mode, so with 4 controllers on your server you have a maximum of 5 disks per controller.
I would definetly stripe across all 10 disks per HVD10 and replicate the same setup on the mirror.