Jim Moyle has recently published a great paper on Windows 7 IOPs – I strongly recommend anyone that wants to understand more about VDI storage gives it a read. With Synergy soon upon us, IOPs is going to a hot topic this week.
Jim’s paper is a great reminder that delivering desktops as a service requires a utility computing mind-set. It’s uneconomic to provision a system to allow for theoretical peak concurrent load – being able to boot 1000’s of desktops in 12 seconds (the time Jim measured to boot a desktop with unconstrained IOPs) would mean designing your infrastructure for 0.014% of the typical working day. Not a practical proposition.
We’re all familiar with how classic utilities address this – ISPs routinely provision their consumer service for a contention ratio > 1, to make the service economically viable – and use traffic shaping, bandwidth and usage constraints to deliver a ‘fair share’ of the ISP’s resources to each consumer. The same model applies to VDI – we need to approach this with the utility model and aim for sharing the limited resources available to ensure adequate capacity is available to each desktop, so that all users get a great experience.
Citrix has experience in this field – desktop virtualization using the shared hosted model (using Remote Desktop Services) has demonstrated that when users share a single OS, the OS must arbitrate to ensure users get a fair share of the CPU resources available. In the VDI world, it’s the hypervisor’s role to arbitrate between user desktops – ensuring each VM (and so desktop) gets a fair share of the CPU and disk I/O capacity available. I’m confident we’ll see hypervisor platforms continue to innovate to improve the ability to shape and limit the resource usage of VMs – enabling increasingly cost-efficient VDI solutions.
For more depth on desktop virtualization storage, I’d recommend this session at synergy: