Design of Disk Systems with EVA-3000 and SecurePath

[Ian Dennison_1]

A year or so ago we implemented a pair of EVA-3000s and SecurePath on our HP-UX site, and I could not find much very information about the “one big disk” approach proposed by the vendor (250GB on 1 disk with SecurePath between the OS and the Array).

I have since been troubleshooting a disk performance monitoring problem from Measureware and Openview and have discovered the following issues. I hope they contribute to the knowledge pool about using large disks and intermediary performance software.

I created a 112GB disk on an EVA-3000, and created 108GB of database space in 16 varying lvol sizes (ranging from 2.6GB to 15 GB – basically a partial copy of our Production Oracle Database). I then backed it up using 1, 2 and 3 LTO-2 Drives over a 2GB SAN to simulate load and measured the load at varying intervals using “glance”. I then deleted and recreated the space using 8 x 14GB disks, and altered them in SecurePath to use the same Fibre Card (no Load Balancing or Round Robin configured in either case) and re-did the backups again.

Results:-
Backing up 1 Physical Disk Drive to 3 LTO-2s incurred 180-200K blks/sec average, 100% disk utilisation and Disk Queues of 14-35.
Backing up 8 Physical Disk Drive to 3 LTO-2s incurred 180-200K blks/sec average, 40-90% disk utilisation and Disk Queues of 0.5-1.0.

In short, using fewer, larger disks on SecurePath will cause your local disk performance statistics to become meaningless. If you wish to get accurate Disk Performance information, better look towards the array-specific tools. We have basically had to turn off our MeasureWare Alerts for our Production Database Disks in Openview, because they are almost always at 100%.

Interesting point to note; backup time was not different using either one or 8 logical disks, so it does not look like performance is an issue when choosing methods of disk grouping.

Share and En

[Bostjan Kosi]

In heavy load environments with more than just one server beeing havy...you can gain some poerformance by creating smaller Vdisks and combining them on the OS level, because you can load ballance the FC hba and also EVA host ports. If you have just one "big" drive all the traffic to that drive goes to only one EVA host port...

B

[Uwe Zessin]

A single host port can handle up to 2048 outstanding I/Os, but the default depth on HP-UX is 8 I/Os per LUN. On the other hand, it makes sense to use at least one more virtual disk that you can assign to the second contoller.