Guest Post by Anjan Srinivas, Director of Product Marketing, Atlantis Computing on how they solve the VDI IOPS Challenge when using ESX/vCenter as the virtualization platform.
Correct IO sizing is always a big challenge for customers adopting VDI. Storage can represent more than 60-70% of the total desktop cost in VDI infrastructure. Get it right and the project is successful – users accept it and desktop cost is low. Get it wrong and the desktop is very expensive and the team does not get to scale beyond the first implementation – or the desktop is inexpensive and users reject it for performance reasons. This challenge becomes amplified when IT tries to use VDI as a means to simplify migration from Windows XP to Windows 7.
The Windows OS was designed with a local and dedicated disk and requires constant access to the hard drive even when it is idle. In addition, the Windows OS will consume as much disk IO or throughput to the hard drive as is available. Windows desktop workloads are write heavy (70-80% writes, 20-30% reads) .Windows 7 OS images are also larger than XP images, forcing enterprises to buy more storage capacity to accommodate larger numbers of users.( a Windows 7 image ranges from 25GB to 50GB or more depending on the components and apps installed)
There is another problem with IOPS when it comes to VDI. All IOPS coming out of virtual desktops are typically treated as “equal” by the hypervisor. This causes a lack of consistent user experience (as user workloads vary). Imagine a user running a zip file compression or running an on-demand virus scan on the same host as the CEO who needs his desktop to work on his board meeting presentation. This can be a recipe for disaster.
So why is IOPS a limiting factor?
Essentially the problem comes down to the physics of a spinning disk. In a traditional hard disk drive there is a spinning platter which is why all disks have an RPM value. Each disk can provide 65-200 IOPS per spindle depending on what type of disk is being used (SATA, SAS). Customers sometimes size for the storage capacity or average IOPS. Both these approaches may result in under sizing the storage.