disk allocation strategy in EVA

[klb]

Hello group, I've got an EVA 5k with 3TB storage. I'll be sharing this space among 3 UNIX servers and am curious about best practices with regard to exactly how the storage is supplied to these boxes.

I intend to use the LVM at the OS level and will basically be building a number of filesystems for RDBMS (cooked) and regular user files on each host.

Currently the size and number of required filesystems is unknown and will certainly grow/change over time. There is also the need to have different raid levels.

Given all the above, I was considering simply creating two 100GB LUNs for each host, one of which would be raid-1 and the other raid-5. Once presented to a host, the LVM would be used to allocate space and build the system from a single volume group containing the two "drives".

Having a few larger Luns as opposed to many smaller ones sounds cleaner and easier to manage, but I could be missing something.

I'm open to all comments and suggestions and am particularly interested in your methods with respect to configuring and presenting space to hosts.

Thanks,

-Kevin

[Jefferson Humber]

Kevin,

How you configure your EVA depends on how you rank the following 3 criteria in order of importance;

a) Resilience
b) Performance
c) Cost

The 'EVA Best Practices' guide can help with the design decisions more once you have the above nailed.

Jeff

[m saravanan]

Kevin,

Attaching best practices Guide. Check out for any latest releases.

It's ok to create two 100GB LUNs. But, if you planned to create a LUN of size more than 150 GB, I'll suggest you to create two 0r three smaller LUNs. Otherwise tune the kernel parameter "scsi_max_qdepth" on the hp-ux box. The default value is 8.

Regds,
Saravanan

[Pat Obrien_1]

I don't understand why you add raid1 and raid5 to the same lvm group. You also mention utilizing lvm, though the 5k needs securepath for most os, and if you are thinking vcs4X, I heard this is on hold.

You environemnt will definately be different than mine, though I create 2-100gb luns and present via securepath to each vg and then host stripe across these luns for all data and oracle data files with 1 exception and theat is redo logs are 5gb striped in similiar manner though the vxfs block size is 1024 instead of default 8k like other fs.

stripe width for me is 256k. I have tested 4 and 8 member stripes in a caned test. the 8 did 20 minutes faster on the 8 hour test which I deem not worth the admin time to configure and manage. Also tested 64, 128, 512, 1024k stripe sizes to determine sweet spot at 256k.

But then you may be doing things completely different. On a good day my 5k run 350-450MB/s and have peaked at 530MB/s per EVA on a server with 3 eva5k and 8 HBA's. Still testing eva8K but have already found 603MB/s with 2 luns.

[klb]

Thank you for the responses.

Some info I left out:
1 disk group for all eva disks.
protection level set to 2

Secure path is in use, as it is required. My mention of using the LVM was simply to let you know that I don't intend to build JFS's on the /dev/dsk/c***** devices like I observed an HP tech do once.

If I go with larger LUN's ( 100GB or so ), what would be a good starting point for setting scsi queue depth? I don't want to cause problems with the internal scsi bus using an inappropriate value here.

What are the advantages of OS striping across LUN's from the same EVA? Seems to me that since all data is striped across all disks in the EVA by default ( assuming 1 disk group ), additional use of round robin allocation among Vdisks at the OS level would not provide any increase in performance. Maybe this would help balance the load across HBA's but that would require some monitoring of secure path to ensure each member of the stripe set was indeed still on a seperate path.

Thanks to those that have posted to this thread. I'd like to hear other's tales of how they built their world.

Thanks,

-Kevin

[Pat Obrien_1]

I have talked about changing qlength a few times, but I am not having issues. You will also learn there is a hardware qlength and a software. I do see instantaneous software qlength at or about 250, bu A the users are not complaining, and B overall perf throughput is great, and C Hp support has recommended against for me. Your mileage may vary.

The advantage of host striping in my environment is that every lun is preferered to every other EVA controller. I strip across controllers to have them both work at the same rate verse one doing all the work and the other idle. Just a way to sneak more perf as again I drive mine hard. I know others whose perf meter does not even move, though I have the need for speed.

[Ken Grabowski]

Your EVA performance is best with larger vdisk and fewer of them. Using LVM lets you add more space at a later time if you need to. But with RDBM you can usually can point to a new disk, so that is optional.

Put as many spindles into a disk group as possible. The more spindles the faster your performance. In the EVA, strip your disks in a disk group across the drive enclosures. Each enclosure has two controllers. The More enclosures used in the disk group the more controllers used, and the greater the speed.

Performance wise, there is no difference between RAID0, RAID1, and RAID5 in an EVA. HP claims it, and I have verified it for myself. RAID level only effects protection level.

One other thing, make a separate disk group for the RDBMS log files. That way your log output is not competing with you data tables.