I was assisting one of my local team members the other day with sizing a VM for Microsoft SQL. I usually always fall back to this guide from VMware. So I started out with the basic seperation of Data and Logs and TempDB.
Make it look like this:
VM Disk Layout
LSI SCSI Adapter
C: – Windows
Paravirtual SCSI Adapter
D: – Logs
E: – Data
F: – TempDB
Which is pretty standard. Then someone said, “Why do we need to do that?” I thought for a second or five. Why DO we need to do that? I knew the answer in the old school. Certain raid types were awesomer at the types of data written by the different parts of the SQL Database. We are in a total post-spindle count world. No Spindles Bro! So what are some reasons to still do it this way for an All Flash Array?
1. Disk Queues
I think of these like torpedo tubes. The more tubes the less people are waiting in line to load torpedoes. You can fire more, so to speak. Just make sure the array on the other end is able to keep up. Having 30 queues all going to one 2 Gbps Fiber Channel port would be no good. See number 3 for paths.
2. Logical Separation and OCD compliance (if using RDMs)
Don’t argue with the DBA. Just do it. If something horrifically bad happens the logs and data will be in different logical containers. So maybe that bad thing happens to one or the other, not both. I am not a proponent of RDM’s. SO much more to manage. If you can’t win or don’t want to fight that fight at least with RDM’s you will be able to label the LUN on the array “SQLSERVER10 Logs D” so you know the LUN matches to something in Windows. This also makes writing snapshot scripts much easier.
Each Datastore or RDM has its own paths, if you are using Round Robin (recommended for Pure Flash Array) more IO on more paths equals better usage of the iSCSI or FC interconnects. If you put it all on one LUN, you only get those queues (see #1) and those paths. Remember do what you can to limit waiting.
Am I going down the right path? How does this make it easier? Are there other reasons to separate the logs and data for a database other than making sure the Raid 10 flux capacitor is set correctly for 8k sequential writes? I don’t want to worry about that anymore. Pretty sure plenty other VM Admins and DBA’s don’t either.
For me a good exercise in questioning why I did things one way and if I should still do them this way now.
2 thoughts on “No Spindles Bro”
I would also supply #4 backups / restores and #5 replication.
Do you think we could combine #4 and #5 into “Data Protection” and add snapshots/clones too?