posts - 219, comments - 411, trackbacks - 27

My Links

Advertisement

News

Follow billgraziano on Twitter

Article Categories

Archives

Post Categories

Consulting

SQL Server

RAID 10 vs. RAID 5 Performance

One of my clients is in the process of standardizing all their servers on RAID 10.  We discovered one of their smaller servers had been delivered with RAID 5 already configured.  I took the opportunity to run some tests on the drives prior to reconfiguring them into RAID 10.  My tests weren't very scientific.  I just used SQLIO.EXE to generate random 8KB reads and writes.

 

image

 

Reads were just a little bit better but writes were almost twice as fast.  This makes sense as the array isn't spending time computing the parity information.  Some notes on my setup:

  • There are FOUR drives in the array.
  • Each drive was a 146GB 15K SAS drive with 8MB of cache.
  • The test was performed with a 5GB file.  I'd prefer a larger file but my time was limited and the RAID 5 tests were already completed.
  • The controller was one of the low-end PERC controllers from Dell.
  • 64KB random writes showed a similar doubling of IOPS.  I didn't test sequential writes.

Print | posted on Monday, June 18, 2007 10:45 AM |

Feedback

Gravatar

# re: RAID 10 vs. RAID 5 Performance

Hi Bill,

When you say "four" drives, do you mean four, total, or four active? I'm specifically wondering about the RAID 10 config -- did you have only 2 active drives?

By the way, you should read this post from Linchi:

http://sqlblog.com/blogs/linchi_shea/archive/2007/02/07/is-raid-5-really-that-bad.aspx

6/18/2007 4:01 PM | Adam Machanic
Gravatar

# re: RAID 10 vs. RAID 5 Performance

Adam,

That's four total drives -- all are active. In the RAID 10 it was two mirrored pairs.

Interesting post by Linchi. It looks like I found different results. My testing was on the exact same hardware. We just deleted the RAID 5 array and replaced it with a RAID 10 array.

-Bill
6/18/2007 5:54 PM | Bill Graziano
Gravatar

# re: RAID 10 vs. RAID 5 Performance

I'm shocked that the reads were faster with RAID 10; shouldn't RAID 5 be able to read from all four spindles at once?

6/19/2007 3:25 PM | Adam Machanic
Gravatar

# re: RAID 10 vs. RAID 5 Performance

Adam; isn't Raid5 slower because of parity checking? That is, it has to read the data, and the parity, and calculate the read data is correct. Whereas with Raid10 just has to read from whichever mirror is more efficient getting the data (closest head to the track).
6/20/2007 8:08 AM | Mark Harr
Gravatar

# re: RAID 10 vs. RAID 5 Performance

I thought it only checked parity on write? I guess it could check on read and fix any bit errors, but that does seem excessively slow, and I've never heard of it (which doesn't mean it's not correct -- I'm certainly not a RAID expert). Do you have a good link describing this?

6/20/2007 8:55 AM | Adam Machanic
Gravatar

# re: RAID 10 vs. RAID 5 Performance

Bill,

Your observations on random IO write performance are consistent with the presentation EMC gave at our March local SQL Users Group meeting (http://www.stlssug.org/docs/Best_Practices_for_Performance.pdf). In addition to the parity calculation, RAID-5 requires 4 operations per random write request because it it’s essentially a rewrite of data on 2 different disks. RAID-5 might fare better in a sequential read/write tests but performance will vary depending on the vendor implementation.

I think the main point of Linchi’s article is that we shouldn’t be too quick to dismiss RAID-5 because performance can vary substantially on different hardware configurations. The nature of the actual application workload (read/write, random/sequential) are major factors in the optimal configuration. I just wish we always had the time and resources to thoroughly test performance.
6/22/2007 6:50 AM | Dan Guzman
Gravatar

# re: RAID 10 vs. RAID 5 Performance

Linchi's tests were run on a decent-quality SAN. BIll tested an ordinary disk controller. I am not surprised to see a major difference in results. When I stress test an IO subsystem, I try and use a file size that is several multiples of the cache size. That is pretty easy on a standard RAID controller. On a SAN, not so much. One effect I see is that as the SAN or even a single RAID set on a SAN approaches saturation, the performance differences beween RAID-5 and RAID-10 become much more apparent.
6/28/2007 3:24 PM | Geoff Hiten
Gravatar

# re: RAID 10 vs. RAID 5 Performance

you might be interested in the theoretical figures
http://sqlblogcasts.com/blogs/grumpyolddba/archive/2007/01/26/how-fast-is-a-disk.aspx
7/9/2007 4:56 AM | colin leversuch-roberts
Gravatar

# re: RAID 10 vs. RAID 5 Performance

i want to know more about raid
7/14/2007 12:02 AM | jatinder singh
Comments have been closed on this topic.

Powered by:
Powered By Subtext Powered By ASP.NET