In my youth, I was offered credit cards. I thought they were great, allowing me lots of instant retail gratification, I was on cloud 9 on the high street…….then I had to repay them. My cloud evaporated and I came back down to earth with a bump right on my wallet !
The principal behind thinly provisioned disks on VMWare ESX storage LUNs would appear to be along the same lines as that of credit cards. Utilising more than you actually have at your disposal.
With credit cards the impact is that later down the line you have to repay the money you spent that you don’t actually have (ouch !). With thinly provisioned disks the pain is that of not being able to reclaim free disk space without a lot of work.
With normal *fat* disk allocation when you create a volume all the space is allocated at creation time. So if you have a 100GB storage LUN and you create a 50GB volume on it, that 50GB is immediately deducted from the 100GB, leaving you with only 50GB free space on the storage LUN.
But with thin provisioned disks, if you allocate a 50GB *thin* volume, the space is only deducted from the storage LUN as it gets used. So if you only write 20GB of files to the 50GB thin volume, the storage LUN will report 80GB of free space out of the total 100GB.
But here’s the kicker, if you then delete 10GB of those files, the free space remains at only 80GB, the space does not get returned back to the storage LUN. This is because Windows/Linux does not actually delete files, it simply marks those blocks as being avilable for use in the file table. As the blocks do not actually get cleared, VMWare does not pick up on this and remains ignorant to the free’d up filespace.
In order to shrink the disk and reclaim the space, you have to actually clear the physical blocks so that they are actually empty. You can do this using the free tool Sdelete from Sysinternals. You use the -c switch (which tells it to zero out the blocks). This is a rather I/O demanding task best done out of peak useage hours.
Once sdelete has completed you will need to migrate the VM files to another datastore for VMWare to re-read the free blocks and give back the cleared up space. You may then have to re-migrate the VM files back to where you actually want them to run from if your fussy about the location of your VM’s on your datastores.
I’m not saying thin disk provisioning has no place, it’s great for R&D, labs and proof of concept type setups where you will be setting up and ripping down and don’t really care too much about long term storage levels. But for production systems, the administration overhead is just too great for my liking.
disk squish
