Q&A with Mark Davis of Virsto Software
VSM: Are there unique problems that virtual desktops create in regards to storage?
MD: There are unique problems in virtual computing environments in general, but these problems actually become worse in virtual desktop environments. The problem is because you have multiple independent virtual machines (VMs) generating their own I/O streams on each host, you have a much more random & more write-intensive I/O pattern in virtual computing than you had in physical server-based computing. The more VM images, the more randomness. With this type of I/O pattern, today’s disk technologies tend to perform at their worst. The storage performance tends to slow down by at least 30% and can be as much as 200%. This means that, to meet your IOPS requirements, you’re going to have to have a larger (i.e. more expensive) storage configuration than you would have had to meet those same requirements in a physical server environment.
VSM: How are these issues the same and/or different from the storage issues faced in virtual server environments?
MD: The problem – storage performance slowdown – is the same in theory; it merely tends to be worse in virtual desktop environments. In virtual desktop environments, although each image is smaller, you tend to have a lot more images, say 50 – 70 instead of the 8 – 12 you tend to have in virtual server environments. This means that the I/O load in VDI environments is more random and more write-intensive than it tends to be in virtual server environments. So disk technologies tend to perform even worse in these environments from a performance point of view.
VSM: How do the storage performance requirements differ in virtual desktops and virtual servers, and why?
MD: Storage performance demands are different in virtual desktop environments in three ways: first, the I/O pattern tends to be even more random and more write-intensive than in virtual server environments, leading to an even greater slowdown with legacy storage technologies. Second, you tend to have much greater variability between peak and average IOPS requirements, due primarily to phenomena like boot and login storms. Third, you are dealing with thousands of desktops and storage objects in virtual desktop environments, whereas in most virtual server environments you are dealing with tens or possibly hundreds. This has implications not only for the scalability of the user interface needed to effectively manage these environments, but also on the scalability of the snapshot technology. For example, if you want to update your golden master and then refresh your desktops, and you have 3,000 desktops, how are you going to do that if your snapshot technology only supports a maximum of 512 snapshots?
VSM: What are the issues around current solutions for virtual desktops?
MD: The typical response to the storage performance slowdown has been to throw hardware at the problem. If you need more IOPS and you’re getting fewer IOPS/spindle then you add more spindles. Or you can add faster spindles. Or you can use solid state storage (SSD). But there are some caveats here. If your performance slowdown was 30%, then in making that up by adding more spindles or faster, more expensive spindles, your costs are going to go up significantly. A common scenario is that these unexpected storage costs climb so high that the cost per desktop becomes prohibitive and the virtual desktop project is shelved.