Take the 2-minute tour ×
Server Fault is a question and answer site for professional system and network administrators. It's 100% free, no registration required.

I have very little experience with SANs, so please excuse this newbie.

One of our production sites has an HP EVA8000 SAN with a bunch of disks (forget exactly how much). it has been configured into a bunch of raid5's and the SQL server uses one RAID volume for data and another for the Log.

So far so good. Or so I thought

Today i ran some speed tests. Using rdfc i created a 100 MB file on the main data disk while the system was heavily used. This showed between 5 and 20 MB/s throughput for write, which, i thought was a bit on the low side. I then took down the application for maintenance, and ran the test again. This time on a quiet system I got 50MB/s write performance.

Is that not a bit on the low side? I mean, I get the same from the SATA drive in my desktop.

I then tried writing 500MB to one drive (the data drive), while writing 500MB to another drive (the log drive). Those two writes interferred with each other. Is that a sign of something wrongly setup?

I don't know what I had expected, but I'm currently reviewing the system to try and locate bottlenecks.

So, i guess my question would be. What kind of performance would you expect (read/write) from a (probably) expensive SAN using Fiber connectors?

(No, its not the little netgear thingy; its this) http://h10010.www1.hp.com/wwpc/ca/en/sm/WF05a/12135568-12135820-12135820-12208992-12208992-12236532.html

share|improve this question
1  
In one of your comments, you say that you were actually getting 50 Megabytes per second, which is normally written as 50MB/sec. You've consistently used Mb/s above, which is Megabits per second. Can you clarify what you're actually getting? –  Daniel Lawson Jun 22 '09 at 22:22
1  
Sorry, my bad. Have changed it now. –  Soraz Jun 23 '09 at 4:32

6 Answers 6

up vote 4 down vote accepted

We have an EVA6100 so I've done tests on similar hardware. The 8000 is, IIRC, an older system than the 6100.

First off, your writes to the DB volume and Log volume are very likely contending for writes as it sounds like they're on the same Disk Group inside the EVA. A disk group is a grouping of disks that you create LUNs on, and each LUN is a disk presented to one (or more) servers. It is not at all uncommon to house many, many LUNs in a single disk-group. If you only have one disk-group, then the benefits of separating your log I/O from your DB I/O are somewhat reduced.

Second, if that DG is running into true I/O contention, then it'll degrade performance. This is something that's very hard to check from a consumer point of view, it's the kind of thing the SAN maintainers need to be aware of. Ask them if they know what the average controller CPU load is, as that'll significantly affect your write performance.

Third, know what your I/O mix really is an benchmark for that, not just file-copy speed. If your DB volume I/O is 80/20 read/write, then you really want to check read-path throughput. If it's the other way around, then your write performance is the metric you really need to check.

share|improve this answer

By simply writing a large file you do not measure storage performance. You just measure the performance of writing a large file. You are sadly mistaken if you think you will get anything similar when: writing randomly all over the disk, writing many blocks simultaneously, writing really large or really small blocks. I guarantee you, that you can have from your array max 1 MB/s on one type of workload, and then with another workload/setup you will easily get 800 MB/s (...provided your server can handle that much).

If you want to learn about these, use a proper tool. I've used Intel IOMeter at the time and I learned a lot with it. Probably now there are some better tools. The main parameter to watch out for is the queue size and random-vs-sequential operation. Read about disk mechanics, why the seek time is so unbelievably slow, and how RAID5 functions.

If you want to know how your storage is performing in real life, take a look at PhysicalDisk counters in perfmon.exe during your work day. Then create some load on your actual application and observe what happens.

share|improve this answer
    
Is IOMeter safe to use on the real disk, or should I get a new set of disks setup for this? My PhysicalDisk counters are all over the place. Average Disk Queue Length goes (visually in perfmon) from around 20 to 100 back down to zero (shortly). I would say it averaged perhaps 35. (I never did grasp how to read the scale part, so perhaps 35 is really 3,5, i dunno). –  Soraz Jun 22 '09 at 20:55
    
+1 for this. RAID5 is not the recommended storage for databases, especially not the logs. –  pauska Jun 22 '09 at 21:17
1  
Soraz, IOMeter is not safe for production. First thing it does is the test file, and it tries to fill up your logical disk to 100% percent with it. Obviously not nice. You have to use restart trick to alleviate this. –  kubanczyk Jun 22 '09 at 21:41

Another important thing to check is for proper disk alignment. More than once, I have seen non-aligned LUNs be the source of big performance problems. If the offset on the partition isn't set before formatting, you can see 25% performance hits on IO.

RAID 10 is also a better option, at least for the log LUNs.

Finally, you can adjust your read/write cache ratio on the LUN to 100% write, since SQL Server does a great job of caching for reads.

SQLIO is a good tool for benchmarks, but definitely not for production. It is easy to use, runs sequential and random read/write tests of variable size, and makes for a good baseline to compare to other systems.

share|improve this answer
    
    
That's a great article, thanks. I read that as "Assume no and check alignment." –  Edinor Jul 18 '09 at 3:22

Look in to RAID10. A smart coworker once convinced me that it's better than RAID5 for write intensive databases, and he was right. It is.

share|improve this answer
    
Amen. The only argument against RAID 10 is reduced capacity, which IMO is a BS argument since platters grow at least 20%/year. –  duffbeer703 Jun 23 '09 at 3:34

It doe seem very low, EVA8xxx boxes are pretty quick when you consider they're built for ease of use. We have lots of them for various things and I have a spare 8100 with 112 * 1TB 7.2k disks and I can easily flood a 4Gbps FC connection, even on a RAID5 vDisk - that's about 385MBps.

I think something's up, I'd happily offer to help but it could take a while, you might be better getting HP in if you have a good enough relationship with them.

Let me know how you get on ok.

share|improve this answer
    
Thank you for the reply. You say 385MBps, is that Mega-bit-per-second? If so, does that mean my 50 Megabyte / second roughly matches that? –  Soraz Jun 22 '09 at 20:40
    
After noticing i write Mb/s instead of what i mean MB/s, i take it yours means 385Megabyte per second –  Soraz Jun 23 '09 at 7:10
1  
I did, generally a upper-case/capital 'B' is bytes while a lower case 'b' means bits. –  Chopper3 Jun 23 '09 at 7:41

What kind of disks are in the SAN (fibre channel, sata etc) , that can have a major effect on the speed your getting. Also when you are doing the tests i take it you are copying from the servers directly connected to the SAN?

share|improve this answer
    
Not really sure as to the types of disk, but I my test setup was "Create an XMB random file on the filesystem" from the server it was attached to, or mounted on (what is that called?). Will check types. –  Soraz Jun 22 '09 at 20:37

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.