Storage systems struggle to provide VDI IOPSDate: Jul 31, 2014
Virtual desktop infrastructure is a storage system's worst nightmare, according to Howard Marks, chief scientist at DeepStorage.net.
In a virtual desktop infrastructure (VDI), desktops are hosted as virtual machines (VMs) on a centralized server so they can be accessed, regardless of an end user's location. But that convenience comes at a price. VDI IOPs are very unpredictable, which means it's common for performance to take a hit due to the numerous VMs located throughout an organization.
"We don't have 20 VMs. We have 200 or 2,000 or 10,000 VMs. We have highly variable workloads -- much more variable than a server workload. [For example,] any individual user can decide, 'I can't find that letter to Smith. Let me use the Windows search function against the file server and have it do a full text search for every document of the 2 million documents we have.' [That kind of search] sucks up input/output like crazy," Marks said.
More on VDI IOPS
How to create a VDI IOPS plan for each desktop
IOPS supply and demand can contribute to VDI capacity planning
Understanding IOPS for virtual systems
In addition, Marks explained why it's so difficult to support VDI IOPS at a reasonable price when the frequency of reads and writes varies. There are typically spikes in activity during any given day, and some are more difficult to predict than others. Boot storms due to users logging on in the morning, for example, are anticipated instances of high reads. But latency can be avoided by using solid-state drives as a read cache. At other times of the day, Marks said, it's harder to predict because of the high number of writes.
"Writes are harder than reads if you're a storage system. If it's mirrored, I've got to write to two locations. If it's RAID 5, I need to read back; I need to calculate parody for small reads with RAID 5. It actually gets really complicated. I still need high performance in the middle of the day," he said.
The best approach, Marks said, is to build and plan your environment for the peak amount of IOPS.