Hi
We recently implemented some dynamic storage shelves in our datacentre. These particular shelves break blocks into pages and move them around the storage capabilities in 2MB chunks (based on various metadata) Our MS based virtual estate uses CSVs and each VM has an agent deployed. The CSVs are located in various different volumes, some are dynamically optimised and others not.
We have been noticing large spikes in our T1 flash usage at the start of the backup window across all volumes where T1 (flash) is incorporated into the volumes. As the tiering of the storage first attempts to write data into T1 and move it later, it would indicate that a process, care of the backup solution (DPM 2012R2), is causing the shelf to first write out data to a virtual disk file within the VM which is in turn being handled in T1 first.
We have moved all of our "system state" scratch partitions to a special virtual disk file where T1 is not used, this has helped a little, but each evening from about when the "sync just before backup window" process begins, data usage increases significantly. LUN growth is negligible. The "data usage" I refer to is the shelf's way of tracking "modified blocks" or "churn" it's not necessarily growth.
I'm starting to think that this might be related to the VSS shadow copies on the volume that are being backed up, and from a change perspective, the entire shadow copy is being modified. Even though for system state the "windowsbackup" folder and data is elsewhere (non-T1) the shadowcopy data still attributed to the source volume. Presumably the same principles apply for other "non-system state" data.
Is DPM caching its copies to the DPM server elsewhere before / during the copy - or perhaps this is "churn" inside memory which is being written in pagefil.sys...
I ran ProcMon to get a better view of this, but was plagued by consistent crashing..
Any ideas?
-a