VDI IOPS

By Hank Smith
Posted in Virtualization
On August 19, 2012

IOP/s in a VDI environment revisited

Ed. Note: The following article was originally written by Hank Smith and published in Gotham’s newsletter in April 2010. An update from Hank appears at the end of this article.

What are IOP/s?

Input/Output Operations per Second (IOP/s) is a measurement of total operations per second (including reads and writes), and is a critical factor in any Virtual Desktop Infrastructure (VDI) project. Typically, the faster the drive, the greater the IOP/s. If the IOP/s for the virtual desktops is not calculated correctly, the VDI solution can fail due to spindle count.

History

When VDI first came on the scene there were two major issues: lack of a connection broker and SAN space requirements. Both VMware and Citrix addressed these concerns. VMware uses the View Manager and vComposer technology and Citrix uses the Desktop Delivery Controller and Provisioning Server to resolve those issues. While required storage space has drastically decreased when using virtual desktops, the requirement for dedicated spindles to host the virtual desktops is still a consideration.

VDI IOP/s Examples

Citrix recently released a case study (http://support.citrix.com/article/CTX123684) that provides the following metrics for IOP/s (averages for 3312 virtual desktops):

Max disk read/write ratio

Reads - 17.80%

Writes - 82.20%

IOP/s

Mean IOP/s per desktop - 4.4

Max average IOP/s per desktop - 27.7

Calculating Spindle Count

Determining I/O based disk requirements is dependent on calculating the percentage of Read IOP/s to Write IOP/s while factoring in the overhead required for RAID mirroring or parity calculations. The following formulas and data are utilized in this effort:

(Reads + Writes x (RAID Overhead factor)) / Spindle Speed = Required Disk

RAID overhead factor = 2 for mirroring; 4 for parity

Reads = max read % * Total IOP/s

Writes = max write % * Total IOP/s

Spindle speed for 15K drives = 180

Based on an example environment comprising 100 Citrix Virtual Desktops, the following calculations can be applied to size the number and performance of spindles required as cited through Citrix Case Studies for both Average and Maximum resultant I/O requirements:

For Average I/O requirements:

Total IOP/s = Average IOP/s per Desktop * # of desktops = 4.4 * 100 = 440

Reads = 17.8% * 440 = 78

Writes = 82.2% * 440 = 362

Calculation for RAID 1/0 (mirroring)

(78 + 362 * 2) / 180 = 4.45 Drives

Calculation for RAID 5 (parity)

(78 + 362 * 4) / 180 = 8.47 Drives

For Maximum I/O Requirements:

Total IOP/s = Max IOP/s per Desktop * # of desktops = 27.7 * 100 = 2770

Reads = 17.8% * 2770 = 493

Writes = 82.2% * 2770 = 2277

Calculation for RAID 1/0

(493 + 2277 * 2) / 180 = 33.51 Drives

Calculation for RAID 5

(493 + 2277 * 4) / 180 = 53.3 Drives

Summary

Proper analysis of a proposed VDI solution is critical to the success of the project. Gotham Technology Group has the tools and engineers to provide this critical data and implement the given solution.

IOPS Revisited – August 2012

In the two years since the article above was first published, Gotham has continued to implement VDI solutions and the IOP/s calculations remain a key factor in their success. One item that has changed drastically since this article was originally published is the average IOP/s per desktop. Using tools like LiquidWare Labs Stratusphere FIT or Lakeside SysTrack, we’ve seen IOP/s averaging around 15 per desktop.

For Average I/O requirements:

Total IOP/s = Average IOP/s per Desktop * # of desktops = 15 * 100 = 1500

Reads = 17.8% * 1500 = 267

Writes = 82.2% * 1500 = 1233

Calculation for RAID 1/0 (mirroring)

(267 + 1233 * 2) / 180 = 15.18 Drives

Calculation for RAID 5 (parity)

(267 + 1233 * 4) / 180 = 28.88 Drives

Of course this varies from customer to customer. For example, the IOP/s can be as high as 80 for customers with various security products installed on the virtual desktop.

IOP/s remains a key calculation when architecting any size VDI deployment. You can have the mother of all server deployments with multicore processors, gigs of memory, etc., but if the IOPS are not taken into consideration the VDI deployment could fall flat on its face.

Hank Smith

Hank Smith

Hank is an expert in the design, development, and delivery of cost-effective, high-performance technology solutions. As manager of Gotham’s Virtualization practice, Hank builds motivated, productive teams for Gotham’s large-scale networking and infrastructure engagements, develops implementation standards and methodologies around virtualization technologies, and manages overall design and implementation of multiple infrastructure projects.