HP Surestore Virtual Array 7100

[Simon White_1]

I am currently evaluating the Surestore Virtual Array 7100 with dual fibre controllers. Each controller is connected via fibre to an HBA in a single LXr8500 running Windows 2000 Advanced Server. I have experimented with a variety of different configurations but the maxiumum transfer rate of am seeing is 10MB's. The advertised rates are 90MB's for reading and 30MB's for writing - I am seeing nothing like that. The array is configured with 7x18GB(15K) disks and 8x36GB(10K) disks. I am also using HP Autopath software to take advantage of the dual paths (without this Windows 2000 gets in a mess). Has anyone got any experience with this new 7100 array and what is the best configuration

[Vincent Fleming]

What's your blocksize???

Any array will show low transfer rates with small blocksizes. 10MB/s sounds like a very high number if you're doing 512 byte I/Os - it would translate to 20,000 I/O's per second. That's way fast! Assuming that you're using something more reasonable, like 4K blocks, you're still around 2,500 I/Os per second. Also very good.

Try a larger blocksize. 64K should get you closer to the 90MB/s. Multiple streams should really get you there.

There are industry standard benchmarks out there that you can use to test bandwidth (your 90MB/s) and IOPS (which should be up to 12,000 IOPS.)

Search for iozone and bonnie on the net for example benchmarks that run on windows. There are many more free ones out there.

Have fun

[Simon White_1]

Thanks for the reply. I have since changed the allocation unit's to 64K (they were set to default previously) when formating the partitions but this still hasn't changed the transfer rates - I am still only getting 10MB's. I have reformatted all 4 partitions created on the 2 LUNS with 64K units.

[Vincent Fleming]

So, you're using NTFS; 64K blocksize now. What are you using to generate the load? (ie: which benchmark?)

If you're just using windows explorer to copy a file and watching the transfer rates, that won't perform well. The copy facility waits for each I/O to complete before sending the next one - meaning that it will defeat the cache on the array. You will see much better results using a real benchmark, such as IOZONE or BONNIE.

I will be setting up an Win2K box (an LT 6000) with 2 VA7100's today. I will reproduce your environment if you post more specific information on how you have it set up, and let you know what I see for transfer rates.

[Simon White_1]

I was just using Explorer to copy files around (sometimes it did make use of the cache on the controller though), but now I have downloaded iozone to see if I get different results.

Currently I have the virtual Array configured as follows:
15 x 18GB (15k) disks.
2 x 512MB controllers linked to two HP HBA's
2 LUNS configured each 40GB in size.
Active hot spare configured.
The capacity of the array looks as follows:
Logical - 80GB
Unallocated - 41.73GB
Redundancy - 95.23GB
Active Spare - 33.38
Total - 250.34

HP Autopath is installed and configured to use default load balancing.

The LUNS seen by Windows are configured as basic disks with 2 primary partitions of 20GB on each giving me E,F,G and H drives.

[Vincent Fleming]

An update for you... I was able to produce 20 to 25 MB/s writing with Explorer file copy to a single NTFS LUN. I also downloaded iometer from intel.com, and was able to see 60MB/s. I still haven't hit the 90MB/s, but I suspect that the single-processor box I'm using might be a constraint.

I'm also not using AutoPath, which may be required to get the 90MB/s.

[Simon White_1]

Vincent, thanks for the reply. Can I ask how you had your system configured, LUNS, block size etc. I have since run some tests using iozone and am getting 48MB's for reading. Is there any benefit in using dynamic volumes in W2KAdv and striping across LUNS as this is the recommeneded approach for Unix ?

[Vincent Fleming]

Hi; We actually decided to use iometer from Intel. Iometer knows how to do multiple threads. We used a single 50GB LUN, RAID 0/1, on a VA7100 thru a Brocade 2800 switch in Fabric Login mode. We did not use AutoPath, and were thus using only 1 port (and 1 controller of the 2). The controllers have 512MB of cache each.

Iometer was set for sequential read, 64K blocks. I forget how many threads we used - it was the default. I don't remember if iozone does this or not - I haven't run either one in 2 years or so.

I would expect that if we install AutoPath so we can load balance between the controllers, we'll see the 90MB/s.

As for striping, well, with a VA7100 I don't think it would improve performance. Use the AutoPath for that.

If you're going to install a database on the VA, you shouldn't be concerned with sequential performance. It's the IOPS that you should be really worried about. Both iozone and iometer can generate small block random reads and writes as a test, and see how many I/O's that thing can do.

Good luck

[Vincent Fleming]

I forgot to mention that the filesystem on the single 50GB LUN was created with the default blocksize for NTFS. The VA has 15 36GB 10,000 rpm drives, and 512MB cache. Your unit sounds like it has faster drives.

[Simon White_1]

Thanks for your time Vincent, in trying out the configuration, it was appreciated. I have since tried different blocksizes, Autopath on and off, striped LUNS, single and multiple LUNS and I receive similar results to yourself 56MB's reads and 30MB's writes. After speaking to HP it seems that the 90MB's result was from a Unix box.