Thin or thick (continued), the empirical new conclusion!

Once wrote a blog about VMware Virtual disk format: "Thin or thick?" Virtual disk format selection problem, introduced thin and 2 kinds of thick format virtual disk features, also mentioned in the article vi3.x time, Deployment (deploy) or Clone (clone) a virtual machine, can not choose the format, can only be Eagerzeroedthick format, This is a disadvantage of vi3.x, and this shortcoming was overcome in 4.0, now we vSphere4.0 when from the template deployment or cloning a virtual machine, can choose to be thin or thick. This improvement seems to imply that a thin deploy from a template is better than (or faster than) deploy a thick virtual machine.

According to common sense, shouldn't this be so? -Because the thin format needs to copy the vmdk file size is relatively small, so deploy speed should also be relatively fast. But I have a question hanging on my mind: how much faster? Is this a determining factor that we should all choose thin format? So, I designed the following experiments.

The lab server is an HP ProLiant bl460c Blade server with 2 4 core CPUs and 12GB of memory, 2 mirrored 146GB 10K to SAS drives, and shared storage is NetApp fas3020c, with software iSCSI connectivity, The MPIO is used in the design of the storage network. The VMFS LUNs for shared storage are experiment-specific, with no other load on it, and the blade server is only an experimental VM, with no additional load. The reason why such a specific description of experimental equipment, is to express a clear meaning: server, network, storage performance is sufficient, does not cause performance bottlenecks.

I installed a Windows XP VM in a clean way from scratch, named Vm-thin, configure memory 1GB, memory 256MB, configure hard disk 8gb,thin format, turn off display space 2.02GB, display occupy space when powered up ( Because there are 0.75GB of swap, why 0.75GB? To understand the principle, please look at the humble text "VMware memory allocation". Then clone the exact same VM, named Vm-thick, and convert the hard drive format to thick when replicating. This thick is actually zeroedthick. Vm-thick off the display of space is 8GB, boot is the display space 8.75GB.

Then I move the 2 VMs in the same state, through storage VMotion, migrate them from the FAS3020 to the local disk, and then migrate back to record the time separately. I then converted the 2 VMs to templates, and then I used the template to deploy the new VM and record the time. Finally, I migrated the 2 VMS deployed from the new template to the local disk again from the FAS3020. In the course of these storage vmotion, I chose the same format as source option.

vm-thin (2.02GB +0.75GB swap)	(8GB+0.75GB swap)
5 "	2"
esx local→fas Storage VMotion	3 ' 02 "	1"
deploy a new VM	2 "	1"
fas→esx local Storage VMotion	5 ' 45 "

The results of the experiment were surprising: the performance of the thick format was much better than the thin format. The exact opposite of what I thought before!

This experiment obtains the following conclusion:

(1) The thick format is faster than the thin format either when deploying a new virtual machine from a template or storage vmotion 40%-50%

(2) Good network-shared storage performance far superior to local disk

This experiment also has an interesting finding that when storage vmotion, the new swap file size is generated in the destination regardless of memory retention, but is exactly equal to the memory size. So before the Svmotion disk footprint or 2.77GB and 8.75gb,svmotion after the end of the 3.02GB and 9GB, browse datastore found that the difference is the size of swap caused by the change. When the shutdown is restarted, the swap is regenerated, and the size of the 0.75GB is restored.

Because the thin format also has a negative performance impact than the thick format when used by a VM. Because the disk in the thin format is dynamically expanding, a vmdk file of several gigabytes is not generated once on disk, so it does not occupy contiguous disk space as a thick disk, so when accessing a disk in the thin format, The disk IO performance is bound to be affected by the longer addressing time that the head moves between discontinuous disks.

To sum up, whether in deployment or application, thin format performance is not as good as thick, strongly recommend that you use thick format virtual disk.

Search, to the last found that the original "default" is the best.