George Crump, Senior Analyst

The first data reduction step is thin provisioning, which only allocates storage as needed. The goal is to exploit the fact that most desktops/laptops are using less than 50% of their local storage capacity, and to only allocate storage to their new virtual counterparts as needed.

The second step in storage reduction is to use a master image, often called a clone or “writeable snapshot”, and build virtual desktops from those. This not only saves the capacity of each machine but also eliminates the high level of redundancy between desktops, like operating systems and applications. Since the only data that would need to be stored is user data and customization of user settings, the potential savings could be as much as 90% when these techniques are combined.

For example if you were virtualizing 1,000 users, each with 200GB hard drives, you would need 200TB of storage without thin provisioning and use of snapshots and clones. With these storage optimization approaches, you would only need 100TB and probably less than that. The more desktops that are virtualized the greater the savings.

The challenge with this approach is that hypervisors like Hyper-V and VMware vSphere have significant performance issues when implementing and supporting thin provisioned volumes, snapshots and clones. Thin provisioned volumes have to be allocated as writes are inbound to the storage system from each virtual desktop, requiring effort on the hypervisor, which has to ‘zero out’ a block of storage each time a new write is defined. Remember, it’s not just one virtual desktop that’s using thin provisioning, nor is it a virtualized server environment that has to deal with a few dozen VMs. All of this on-the-fly extension of the volume activity is occurring nearly simultaneously across thousands of volumes assigned to desktop VMs.

In similar fashion cloned volumes can suffer performance issues. The more that the ratio of dependent virtual desktops increases and the more that they differ from the original golden master, the more data the hypervisor has to manage. Clones are largely dependent on the hypervisor’s ability to deliver high performing snapshot copies.

A clone is basically a writable snapshot. As was the case with thin provisioning, a clone requires extra work on the part of the hypervisor. Blocks of a virtual volume that have snapshots applied to them must be tracked for changes, along with the original blocks. Clones make matters worse by presenting the snapshots as “writable” which means even more blocks to track and more work for the hypervisor. This additional work translates into additional cycle times, which slows down performance for the end user.

The typical recommendation to solve these performance issues is to use fixed or fully allocated volumes. While fixed volumes perform well they do so without the benefit of capacity optimization. In the example above, the choice to use fixed disks to support production performance requires using 200 TB of storage. Even on mid-range arrays at $5 to $15 per GB the cost per virtual desktop becomes significantly encumbered. The ROI of the VDI initiative is once again put in jeopardy as large amounts of excess capacity has to be purchased, but goes largely unused. There is also a major provisioning problem with this approach. When a volume is created it must be zeroed out and prepped for use. Provisioning a thousand virtual desktops with fixed volumes can take days.

Another option is to use a higher-end enterprise storage system that can manage the thin provisioning and cloning processes for the hypervisor. The drawback with this approach as stated above is the cost, which may be even more expensive than the fixed volume approach detailed above, even though the approach enjoys storage efficiency, given the premium pricing of Tier I arrays – often $25 to $31 per GB.

VDI Storage Efficiency Solution

At the heart of the VDI storage efficiency problem is the hypervisor’s inability to maintain performance while supporting techniques like thin provisioning and cloning. The tasks for these processes are similar in that they rely on the file system’s ability to track and manage blocks and sub-blocks of data. Consequently, most of the hypervisors are not able to support them at a high performance level, while maintaining all of their other tasks.

An alternative may be to simply add better provisioning, snapshot and cloning intelligence, instead of replacing all of the hypervisor's storage intelligence with an enterprise array. This is what Virsto Software has done with their new approach to solving the storage challenges of VDI with a new virtual machine storage object, called a vDisk.

A vDisk leverages proven journaling technology similar to what an enterprise database uses. The Virsto Log space is an optimized area of disk that queues up writes so that random write performance is sequentialized. This journalling technique extends performance advantages to thin provisioning, snapshots and cloning since the log area allows the hypervisor to respond quickly to requests to and from those areas. All three of these tasks are write-performance sensitive and the log space improves write responsiveness. Benchmarking data show the performance improvements with this approach are surprising – 2-5x faster than other virtual drive approaches. With this type of technology in place, the IT Manager now has the ability to use less expensive, mid-range storage, leveraging the hypervisor to manage that storage while using Virsto vDisks to gain the high performing provisioning and cloning functions.

With new software designed to optimize storage for VDI, the storage efficiency trade off is balanced. Cost effective storage can be efficiently deployed since hard allocation is not needed. Also, an almost infinite number of clones can be made to reduce the storage footprint even further by eliminating the redundancy between virtual desktops, representing a 90% reduction in storage footprint from the original desktop hard drive capacity. The connecting storage platform can now leverage mid-range arrays at $5 to $15 per GB and do so with great efficiency and performance, further driving down the costs. The net effect is that the costs of the VDI project come back in line, and in fact may actually be more economical than the original local storage.

A log based approach has performance advantages that go well beyond making sure that thin provisioning, snapshots and clones work effectively. It also provides broad based performance improvements throughout the VDI storage infrastructure. Storage Switzerland will cover these specific performance improvements in part two of this series.

– Efficiency vs. Performance – Part I

Virsto is a client of Storage Switzerland