MSA/SAN/Raid Performance

[Chris Geier]

I have been looking for quite some time for information on trying to accuratly guage system/IO performance as it realtes to disk I/O. Trying to estimate performance differences between raid levels, drive speed etc. In addition to this judging what system I/O should be and measure it against what we are seeing with testing tools. Currently I believe we are getting HORRIBLE performance from our MSA1000 and I do not believe it is the product or the like but possibly how we have it configured. Does anyone have any REALLY GOOD information or links that could help me help myself?

[Vincent Fleming]

Ah... you've asked the age-old question of how to accurately measure performance...

No can do.

There are a variety of (bad) benchmarks out there - you may find one you like, but IMHO, they all have their quirks, and ar therefore virtually useless.

Look at iozone (www.iozone.org), bonnie (http://www.coker.com.au/bonnie++/), TIOBench (http://sourceforge.net/projects/tiobench), and iometer (www.iometer.org).

I think iometer is one of the better ones, but if I remember correctly, it only works on Windows, Solaris, and Linux. You would have to port it yourself.

The most important thing to stress here is that you run the benchmark again and again, each time changing one (and only one) setting on the system or disk array.

Don't forget that you want something representative of your workload. Generally, the random I/O benchmarks are too random (real apps aren't really all that random; they tend to do I/O in working sets that appear more random than they really are).

Don't forget that filesystem state can be a significant factor as well - a full filesystem will tend to be slower than a brand-new empty one. I say "tend to be" because there are I/O loads and FS structures that can produce anomolies.

Good luck,

Vince

[Chris Geier]

Crud,

I was hoping there was something out there to help me out more with this. I have played with IOmeter but have had issues with such huge differences with results based on configuration settings within the tool. I have found it very difficult to adequately estimate what a production load would look like as far as all the difference read write sizes you put in the options. I have been trying to find good examples of that as well but have not been able to either. My goal in all this is to provide real guidance in how much a performance affect there is by changing raid levels, numbers of disks, RPM speed of disks, and maybe even eventually stripe size and block size. So far the number I am getting back in my load tests for the current implementation are horrible. I have managed to pick apart a few places and come up with this failed or not. See attached

[Randy Weis]

This is a really interesting topic. I got an HP restricted white paper on a comparison of the various HP disk arrays (VA, EVA, XP), but unfortunately not the MSA.
Your textfile breakdown was interesting, too.
Several factors are important here.
1. Application-specific performance characteristics: what apps are you running on this MSA? OLTP database data files, or logs? ERP? Streaming?
2. Processor and server I/O bus architecture: what platform are you on? How many cores? What kind of bus, and is the segment shared?
3. OS: which one are you using?
4. Size and number of disk: what is your array setup - 72 GB or 146 GB; 10 k or 15 k; ultra3, ultra 160 or utlra320?
5. Cache: how much have you set up? I think the MSA goes up to 512 MB now. Are you using Write cache? What is hurting the most, reads or writes?
6. Controller: are you SCSI connected to the MSA or FC? What speed HBA and swtich are you using?

We did some performance testing with a NAS200s, using an MSA30 (Ultra320 shelf for MSA) and a FC connection to an EVA 5000 (2 Gb). The Ultra320 connection blew the doors off the EVA5000. A single app, unshared U320 connection is much faster than a shared environment running at 1 or 2 Gbs.

I'd be interested to here more.

[Chris Geier]

We are running SQL2000 for a database they are trying to scale to 5M rows. we have a mix of drive speeds depending on the array function. Some of that we need to change to have it all be 15K drives. Mostly we are using R1+0. Servers are DL580 quad proc 8GB memory. When we are running load tests the normal resources are not taxed. For example peak processor utilization is at 40% always above 4GB Memory free etc. Bottle neck is at the disk IO level. Really high table scans and high disk queue length.

[Chris Geier]

oops more to answer

1. Cache: how much have you set up? I think the MSA goes up to 512 MB now. Are you using Write cache? What is hurting the most, reads or writes? We are have 256. set at 50/50 Writes appear to be killing us.

6. Controller: are you SCSI connected to the MSA or FC? What speed HBA and swtich are you using? FCA2214 HBA, MSA SAN Switch 2/8


Unfortunately at the moment the DB itself is on the 10K drives RAID 5. So I believe we can improve performance pretty dramatically by going 15K Raid 1+0

[John Kufrovich]

Chris,

I am a member of the MSA team.

Can you provide some information.

HBA model.
How many servers are attached to the MSA. How many Luns presented to each server.
How many HBAs in each server.

Can you provide the tech_support from the MSA cli.

Regards,

[Chris Geier]

ï ± MSA1000 SAN Starter Kit
ï ± MSA1000 Controller w/256MB Cache
ï ± MSA SAN Switch 2/8
ï ± (2) FCA2214 HBAâ s
ï ± Fibre Cables
ï ± HP StorageWorks MSA1000 High Availability Option
ï ± MSA1000 Controller w/256MB Cache
ï ± MSA SAN Switch 2/8
ï ± (2) FCA2214 HBAâ s
ï ± Fibre Cables
ï ± Two 256MB Cache Upgrades (1 per controller)
ï ± HP StorageWorks MSA30 Single Bus Enclosure
ï ± Six 18.2GB Pluggable Ultra320 SCSI 15K RPM Universal Hard Drives
ï ± Six 36.4GB Pluggable Ultra320 SCSI 15K RPM Universal Hard Drives
ï ± Seventeen 146.8GB Plu

[John Kufrovich]

Thanks,

So it appears you have two DL580 attached to the MSA1000, correct?

In your attachment, you mention RAID 5 and RAID 1+0. How many physical disk make up this LUNs. What I looking for is how is your MSA carved.

[Chris Geier]

maybe this will help.