by George Crump, Storage Switzerland

VDI (Virtual Desktop Infrastructure) implementation projects are going to be priorities for many IT Managers in 2013 and a key concern will be end-user acceptance. If the users don't embrace their virtual desktops they won't use them and the project is doomed to failure. The key to acceptance is to provide users with an environment that feels the same, performs better and is more reliable than their current stand-alone system. The storage system bears most of the responsibility in delivering that experience.

IT managers who want to capitalize on the opportunity that the virtual desktop environment can focus on two key capabilities when they evaluate storage system vendors. The first is being able to deliver the raw performance that the virtual desktop architecture needs and the second is doing so in the most cost effective way possible. These are two capabilities that are traditionally at odds with each other and not always well-reflected in benchmark testing.

For most organizations the number-one priority for gaining user acceptance is to keep the virtual desktop experience as similar to the physical desktop as possible. Typically, this will mean using persistent desktops, a VDI implementation in which each user's desktop is a stand-alone element in the virtual environment for which they can customize settings and add their own applications just like they could on their physical desktop.

The problem with persistent desktops is that a unique image is created for each desktop or user, which can add up to thousands of images for larger VDI populations. Obviously, allocating storage for thousands of virtual desktops is a high price to pay for maintaining a positive user experience.

In an effort to reduce the amount of storage required for all of these images, virtualized environments have incorporate features such as thin provisioning and linked clones. The goal is to have the storage system deliver a VDI environment that's built from just a few thinly provisioned ‘golden' VDI images, which are then cloned for each user.

As users customize their clones, only the differences between the golden image and the users' VDIs need to be stored. The result is a significant reduction in the total amount of storage required, lowering its overall cost. Also, the small number of golden images allows for much of the VDI read traffic to be served from a flash-based tier or cache.

When a write occurs from a thinly provisioned, cloned virtual desktop more has to happen then just the operation to write that data object. The volume needs to have additional space allocated to it (one write operation), the metadata table that tracks unique branches of the cloned volume has to be updated (another write operation) and some sort of parity data needs to be written, depending on the RAID protection in place. Then, finally, the data object is written. This entire process has to happen with each data change no matter how small.

Herein lays the tradeoff in using these features. While reducing the amount of space required for the VDI images, thin provisioning and cloning increase the demand for high write performance in the storage system. This presents a significant opportunity for storage system vendors who can address these new performance requirements.

Many storage systems that use a mix of flash memory and hard disk technology don't use the higher performing flash for writes; they use it for actively reading data. While these storage systems have storage controllers designed to handle high read loads, the increased write activity generated by thin provisioning and cloning is still going to relatively slow hard disk drives. Because this type of I/O traffic is highly random, the hard drives are constantly "thrashing about". Basically the controller sits idle while it waits for the hard disk to rotate into position to complete each write command. Even systems with an SSD tier or cache may have problems providing adequate performance because they too don't leverage the high speed flash for write traffic.

Due to the high level of thin provisioning and cloning, plus the fact that once a desktop is created a large part of its I/O is write traffic, many cached or tiered systems do not perform well in real-world VDI environments and can provide misleading VDI Benchmark scores.

The Truth Behind VDI Benchmarks
Most VDI Benchmarks focus primarily on one aspect of the VDI experience, the time it takes to boot a given number of virtual desktops. The problem with using a "boot storm test" is that this important but read-heavy event is only a part of the overall VDI storage challenge. During most of the day desktops are writing data, not reading it. In addition, simple activities such as logging out and application updates are very write-intensive. The capability of a storage system to handle these write activities is not measured by many VDI benchmarking routines.

A second problem with many VDI benchmarking claims is that for their testing configuration they do not use thinly provisioned and cloned volumes. Instead, they use thick volumes in order to show maximum VDI performance.

As discussed above, in order to keep user adoption high and costs low most VDI implementations would preferentially use persistent desktops with thin provisioning and cloning. Be wary of vendors claiming a single device can support over 1000 VDI users. These claims are usually based on the amount of storage that a typical VDI user might need as opposed to the Read/Write IOPS performance they will most likely need.

Trustworthy VDI Performance
A successful VDI project is one that gains end-user acceptance while reducing desktop support costs. The cost of a storage system that can provide thin provisioning, cloning and an adequate sized flash storage area to support the virtual environment could be too high for some enterprises to afford.  And, an additional cost could be incurred with the performance problems that are likely to appear after the initial desktop boot is completed because of the high level of write I/O.

The simplest solution may be to deploy a solid state appliance like Astute Networks ViSX for VDI. These devices are 100% solid state storage to provide high performance on both reads AND writes. This means that boot performance is excellent and performance throughout the day is maintained as well.

With a solid state based solution to the above problems, performance will not be an issue, but cost may still be. Even though it can provide consistent read/write performance throughout the day for a given number of virtual desktops, the cost per desktop of a flash based solution can be significantly higher than a hard drive based system.

However, it's likely in larger VDI environments (400+ users) that flash-based systems are really the only viable alternative to meet the performance requirements which can easily exceed 100 IOPS per user. Fortunately, flash-based systems can also produce efficiencies that bring down that cost in addition to the well-known benefits of using 1/10th the floor space, power and cooling compared to traditional storage systems.

First, the density of virtual desktops per host can be significantly higher with a flash appliance. And, the system is unaffected by the increase in random I/O as the density of virtual machines increases.

Second, the speed of the storage device compensates for the increased demands of thin provisioning and cloning operations run on the hypervisor. These data reduction services can now be used without a performance penalty. This means that the cost of a storage system with a more powerful storage controller and expensive data services like thin provisioning and cloning can be avoided.

Finally, the flash appliance is designed to tap into more of the full potential of solid state-based storage. For example, Astute uses a unique DataPump Engine protocol processor that's designed to specifically accelerate data onto and off of the network and through the appliance to the fast flash storage. This lowers the cost per IOPS compared to other flash-based storage systems.

Most legacy storage systems use traditional networking components and get nowhere near the full potential of flash. In short, the appliance can deliver better performance with the same amount of flash memory space. This leads to further increases in virtual machine density and space efficiency because more clones can be made - resulting in very low cost per VDI user.

Conclusion

VDI benchmark data can be useful but the test itself must be analyzed. Users should look for tests that not only focus on boot performance but also performance throughout the day, and at the end of the day. If systems with a mix of flash and HDD are used then enough flash must be purchased to avoid a cache miss, since these systems rarely have enough disk spindles to provide adequate secondary performance.

A simpler and better performing solution may be to use a solid state appliance like those available from Astute Networks. These allow for consistent, high performance throughout the day at a cost per IOPS that hybrid and traditional storage vendors can't match. Their enablement of the built-in hypervisor capabilities, like thin provisioning, cloning and snapshots, also means that they can be deployed very cost effectively.

>

George Crump is lead analyst of Storage Switzerland, an IT analyst firm focused on the storage and virtualization segments.

• Atlantis Computing and Axept AG Solve the VDI Storage Challenge for Swiss Bank
• Atlantis Computing & Axept AG Solve VDI Storage Challenge for Swiss Bank
• Deep Blue Marine, Inc. Announces New Interim Prize Offer in the "Can You Dig It?" Contest