mirroring: new Storage vMotion method in vSphere 5

Thuleman

Supreme [H]ardness
Joined
Apr 13, 2004
Messages
5,833
Some of you may have run into this in v4. You have a VM that has really busy IO but it needs to be move to another LUN. This could create a problem since v4 Storage vMotion would utilize changed block tracking and move inactive blocks over to the new storage and then wait for active blocks to become inactive to complete the Storage vMotion.

For high IO disks this may just never happen and Storage vMotion would time out. I personally never ran into this issue but I can certainly conceptually see how it would happen if you are running anything that has a lot of IO to disk.

in v5 the problem has been resolved.
What happens now is that there's a block level bitmap which on the host level keeps track of which blocks are active and inactive. When you initiate a Storage vMotion the source will be suspended and resumed for a split second to initiate a mirror driver. The mirror driver then starts writing active blocks to the source and the destination based on the bitmap while the inactive blocks are pulled from the source drive.

So while the VM is in Storage vMotion writes to active blocks occur to source and destination, once the mirror is perfect as all inactive blocks have been pulled from the source again a split second suspend/resume occurs and the source is disconnected and deleted.

Overall this will significantly increase the performance of Storage vMotion of IO intensive VMs, meaning it will significantly decrease the duration of Storage vMotion and prevent time-outs. Fun stuff!
 
Big push for this was due to Storage DRS. Would have gotten real ugly using the old way with SDRS. We do svMotions a lot for migrating between arrays so this makes it even smoother. Would sometimes have issues doing it during the day.
 
I'm curious how this is going to work with Storage Tiering..etc. It seems that you would want to only enable the failover piece of this and not the performance piece as to not interferring with FAST policies..etc.
 
I'm curious how this is going to work with Storage Tiering..etc. It seems that you would want to only enable the failover piece of this and not the performance piece as to not interferring with FAST policies..etc.

Those are good design questions. I see SDRS as having a very slow adoption curve, at least before people click the Fully Automated button. I can see people doing it with FAST VP with multiple pools, but even then it can get messy. It's also going to be real messy right now for those replicating datastores offsite. Move a 300GB VM to another datastore and what happens? You'll replicate a lot of data for no good reason. I know EMC is working to integrate RecoverPoint with this and I'm sure everyone else is doing the same...but until then....
 
Back
Top