Performance - Single VS Multiple Mount Points

• PatLap

  Mr or Mrs. 500

  Points: 524

  More actions

  April 25, 2013 at 8:42 am

  #293035

  Hi,

  I would appreciate any thoughts or comments on the following.

  We have a 3 nodes cluster hosting 12 SQL 2008 R2 instances. Each instance has a minimum of 4 Luns (1 for the instance root, 1 for data, 1 for log and 1 for tempdb).

  Some instances have several data LUN (up to 6 on two of them). All this for a total of 64 LUNs (using Mount points obviously).

  We are soon to migrate to a new Toshiba SAN (using thin provisioning) that will have a 276GB cache with 3 level of drives (12K, 15K and SSD). The SAN will not host only SQL.

  All the data is spread over all spindles with the SAN being intelligent enough to move blocks of data to faster drives depending on requests.

  We want to see if it's possible to reduce the number of LUNs. The high number of Luns is causing some headaches to the DR team when doing DR exercise.

  Since all spindles are used, our first thought is a go ahead.

  The setup is fairly complex. The data, log and tempdb mount points are attached under the instance mount point.

  For exemple you have a drive F: which is the instance LUN and F:\MSSQLDATA\Data1 which is a mount point for data and F:\MSSQLDATA\Log1 the mount point for log, etc

  I'm kind of concerned about impact on the server (node) side of things. I saw from other posts that having multiple LUN creates multiple drive queues in Windows.

  On the other side, we only have a single 10GB NIC card, therefore having one or multiple LUN shouldn't change anything, no? At the end is it just a matter of the NIC card bandwith?

  Would you

  A) keep the status quo

  B) Increase the number of LUN, when possible (like having a pseudo size limit of 500GB or 1TB)

  C) Decrease the number of LUN, when possible but keep the 4 LUN/instances (1 for the instance root, 1 for data, 1 for log and 1 for tempdb)

  D) Decrease the number as much as possible (thinking of going with 1 for the instance root and one for the rest)

  I thought about testing it in our lab, but how would I test it and with which tool?

  How would I use SQLIO in a way to compare throughput from let's say 2 data LUN VS having a single one?

  I did test using 2 test files on 2 different LUNs with 2 threads each and I get roughly the same IOs/sec than 1 LUN with 4 threads.

  Any input would be appreciated.

  Thank you

  Patrick