We’re in the process of testing the deployment of VDI now so I attended a session about preparing for a deployment. As I mentioned in my previous post, I’m not an expert on any subject after attending a one hour session, but I thought I’d share my thoughts.
The most important take away from the session was mentioned near the beginning – for a truly successful deployment you must do a detailed desktop IO analysis. It’s important to have a firm grasp on the amount of desktop IO that your company requires and a detailed analysis is the only way to gather that info. There are “rules of thumb” that can be followed ( 5 IOPs for light users, 10 IOPs for medium users, and 20 IOPs for heavy users), but you could easily end up either over-allocating or under-allocating without knowing the actual numbers. Lakeside software and Liquidware labs were mentioned as vendors who provide software that does such an an analysis, however I’ve never heard of either of them and can provide no information or feedback on their services. There’s a good free VDI calculator available at http://myvirtualcloud.net/?page_id=1076. Once you have a good grasp of the amount of IO you’ll need to support, scaling for the 95% percentile should be your target.
What’s the best way to prepare for VDI on a VNX array? If you conclude from your analysis that your desktop environment is more read intensive, consider hosting it on storage pools that utilize FAST VP with EFD and FastCache. Using VMWare’s View Storage Accelerator and EMC’s VFCache would also be of great benefit. If your environment is more write intensive, focus on increasing the spindle count (even at the expense of wasted capacity). All of the other items mentioned for read intensive environments still apply and would be beneficial.
– If you’re using Windows XP make sure disk alignment is set correctly. If it’s not, you could see up to a 50% penalty in performance.
– Image optimization is very important. Remove all unneeded services.
– Schedule your normal maintenance operations off hours. Running patches and updates during business hours could cause the help desk phones to light up.
– Using NFS vs. block is a wash performance wise. NFS is currently better for scaling up, however, as it allows for up to a 32 node cluster vs. only 8 for block (on linked clones).
– Desktops tend to be write heavy. FAST VP is great!