RAID 10, Crystal Disk Mark Weird score?

  • Hi all,

    so ive been reading Glenn Berrys new book called SQL Server Hardware.

    Ive been using SQLIO to compare my hard drives but decided to try out Crystal Disk Mark as its so much quicker.

    In Glenns book he does a test against a 6 disk RAID 10 and gets 531 Seq Reads and 414 Seq Writes. When i ran the test i got scores of high 200s...

    Apparently he is using SAS 15k disks which is the same as i am. So i thought the scores would be pretty similar. Is it normal for different SANs to give out such big difference in scores?

    Thanks

    S

  • Yes, because SANS are set up in different ways and also have different hardware. Some SANS have 1 disk controller, others have multiples. Also, it depends on how the LUNS are set up and a variety of other things that I only have the barest whisper of knowledge about.

    If you google about SQL Server and SAN, you should get a bunch of articles (some good, some questionable) to read. And just googling SAN alone should get you information on how they are set up and created.

    Brandie Tarvin, MCITP Database AdministratorLiveJournal Blog: http://brandietarvin.livejournal.com/[/url]On LinkedIn!, Google+, and Twitter.Freelance Writer: ShadowrunLatchkeys: Nevermore, Latchkeys: The Bootleg War, and Latchkeys: Roscoes in the Night are now available on Nook and Kindle.

  • Hi Brandie,

    Yeah i figured there would be a ways to improve it, but it seemed a lot higher.. and ive never got mine any where near there.

    thanks for the reply

  • There's a very specific set of articles on SQL Server and LUN setups somewhere that I cannot find or remember the titles of that could help you out. I'm going to call in some reinforcements from another thread and see if anyone there has information that would help you improve your numbers.

    Brandie Tarvin, MCITP Database AdministratorLiveJournal Blog: http://brandietarvin.livejournal.com/[/url]On LinkedIn!, Google+, and Twitter.Freelance Writer: ShadowrunLatchkeys: Nevermore, Latchkeys: The Bootleg War, and Latchkeys: Roscoes in the Night are now available on Nook and Kindle.

  • Search for MPIO - Multi Path Input Output.

  • any info would great ty..

    ive read things like brent ozars chapter in internals and troubleshooting. watched a few vids on configuring you SAN etc..

    This is a brand new SAN (2 controllers, cant be used for multipathing but can do the ozar suggestion of say log file data down controller A and data down controller B with 2 hbas using FC). onto a 6 disk SAS 15k RAID 10 no other data was on the network..

    I just cant seem to get it over 300 seq reads or writes and wondered if there might be something im missing..

    Thanks for any replies and suggestions!

    S

  • hi Greg,

    Yeah that seems like a great to super numbers, but apparently our SAN dosnt handle true multipathing. As i said above (Think i wrote it same time as you replie :D) im trying to help things by using both controllers and both of the Servers HBAs to get a bit more out of it..

    Let me know if im on the right path or if im talking rubbish! knowing either way is good 😀

    S

  • I'm no expert, just learned enough to cover my backside when I run into performance issues.

    Each SAN should have some documentation like this, specific to how to set some things up.

    There used to be some white papers published for SQL 2005 / 2008 for Project Real which also dived into some low level details.

    Can't seem to find them, but you might try searching for them.

    Hope this helps you out.

    Too many irons in the fire today to search around much more.

  • There are all sorts of config options that can have an effect. What's the stripe size? How much cache does the controller (or do the controllers) have, and is it battery backed? If it's battery backed, how have you configured the balance between read cache and write cache space? What are you measuring as reads/writes - something defined in teh disc config, or SQL Server pages, or what? What interfacte do you have from the server to the controller(s)?

    Tom

  • First thanks for all the info... SAN design/performance/monitoring is something i want to look into and start doing as we dont really do it here. Any advice on where to start would be great 🙂

    I ran the test 5 times on each LUNs last night (Only possible outside effect to scew results is data being sent on the same switch to the SAN from Servers (I have a RAID1, RAID 10 with 6 disks and a RAID 10 with 8 disks). Results are below:

    Seq Writes

    Almost identical for all 3 106-107

    Seq Reads

    all over the place.

    For the 8 disk RAID10 i had 2 scores of 155 then 3 around 230-280

    For RAID 1 i had 123-197

    For RAID 10 6 disks 154-276

    Pretty sure i should be seeing a nice increase in performance from 6 disks to 8 disks?

    512 Random

    RAID 1 and RAID 10 6 disk results are pretty much identical! (150-160 Reads on both) (180-185 writes)

    RAID 10 with 8 disks slightly better (202-208 Reads) but again same speeds writes! (180-185)

    4k Random

    Reads on all 3 are 4-5

    Writes for RAID1 are 6-7 where as the RAID 10 for both 6-8 are 14-18

    4k with que depth

    Similar as 4k

    Reads and writes for RAID 1 less than RAID 10 (20-21 r and 4-5 w) but both RAID are the same (30-45 r 48-50 w)

    Any thoughts on the results? going to read up on RAID levels now as im pretty some of the results dont match the norm of what should happen!

  • OK....

    I think i see whats happening.. I think our SAN has just done RAID 1 with 6 and 8 disks not RAID 10!

    Wiki on RAID 1:

    In RAID 1 (mirroring without parity or striping), data is written identically to multiple disks (a "mirrored set"). While any number of disks may be used, many implementations deal with only 2.[citation needed] The array continues to operate as long as at least one drive is functioning. With appropriate operating system support, there can be increased read performance, and only a minimal write performance reduction; implementing RAID 1 with a separate controller for each disk in order to perform simultaneous reads (and writes) is sometimes called multiplexing (or duplexing when there are only 2 disks).

    So my Writes are always pretty close. and my reads go up with each added disks (The difference between 6-8 is very very small) but jumps more from 2-6.

    this sound logical and right to you more experienced people?

  • Verify with your SAN administrator, but it does sound correct. Then again, I've never used Raid 10, so I wouldn't know for sure.

    Or, if you are the admin, verify with your vendor. If they gave you the wrong thing, now's the time to get them to fix it or get a bonus based on the grief they've given you.

    Brandie Tarvin, MCITP Database AdministratorLiveJournal Blog: http://brandietarvin.livejournal.com/[/url]On LinkedIn!, Google+, and Twitter.Freelance Writer: ShadowrunLatchkeys: Nevermore, Latchkeys: The Bootleg War, and Latchkeys: Roscoes in the Night are now available on Nook and Kindle.

  • Hi Brandie,

    Thanks for help and the nod that it seems im going in the right direction.. Phone call is on to SAN vendors, So hoping I can get it confirmed its been done wrong!!

    cheers

    S

  • Glad we could help. Let us know what you find out.

    Brandie Tarvin, MCITP Database AdministratorLiveJournal Blog: http://brandietarvin.livejournal.com/[/url]On LinkedIn!, Google+, and Twitter.Freelance Writer: ShadowrunLatchkeys: Nevermore, Latchkeys: The Bootleg War, and Latchkeys: Roscoes in the Night are now available on Nook and Kindle.

  • stebennettsjb (7/25/2011)


    Hi all,

    Ive been using SQLIO to compare my hard drives but decided to try out Crystal Disk Mark as its so much quicker.

    Go back to SQLIO, particularly using files that exceed the amount of RAM cache available to both the SAN/RAID card and the OS. Crystal Disk Mark may be buffered by the SAN, doesn't provide the level of detail, and isn't as repeatable on most systems where other things happen for short periods of time due to not lasting long enough for momentary changes to average out.

    I generally do each SQLIO test for 1200 seconds when I'm doing serious benchmarking; even at 300 seconds each I've seen the occassional very odd SQLIO result that turned out to be spurious.

Viewing 15 posts - 1 through 15 (of 17 total)

You must be logged in to reply to this topic. Login to reply