In about fifth or sixth grade, teachers start subtracting points if you don’t include units (inches, pounds, meters, etc.) with numeric answers to word problems on math tests. During a recent evaluation of Citrix XenDesktop on the NetApp All-Flash FAS8060 storage platform, I confess that I was guilty of forgetting to add units.
The majority of XenDesktop testbeds I’d configured previously used storage shelves of hard disk drives (HDDs) or hybrid storage solutions with HDDs and a small number of solid state drives (SSDs). Quite frankly I’d become accustomed to read and write latencies of a few milliseconds. When I saw latencies ranging from 450 to 1700, I thought something was wrong and I’d made an error in setting up the test environment. It took me a moment to realize that those latencies were values in microseconds and not milliseconds. This “aha” moment highlights the low latencies of the NetApp all-flash FAS solution and how they can help to speed up desktop virtualization provisioning and boot times for XenDesktop.
The goal of this testing exercise was to evaluate the NetApp All-Flash FAS storage platform with an I/O-intensive pure virtual desktop (VDI) workload. In particular, I wanted to characterize I/O performance for XenDesktop provisioning and boot storm scenarios, comparing the results of using Machine Creation Services (MCS) and Citrix Provisioning Services (PVS) to provision a 2,000-seat VDI farm. I also wanted to examine the impact of storage efficiency features like deduplication. I’ve documented the test configuration, methodology, and test results in a short white paper (available here) that this blog summarizes.
About the testing
In addition to the NetApp FAS8060 storage platform, the testbed (Figure 1) included two network switches and 16 rack-mount servers (14 for VDI desktops and 2 for infrastructure services). Software components included Citrix XenDesktop 7.6, Citrix Provisioning Services (PVS), Microsoft® System Center Virtual Machine Manager 2012 R2 (SCVMM), and Microsoft® Windows® Server 2012 R2 with Hyper-V®. The NetApp system presented multiple iSCSI datastores to the Microsoft Hyper-V hosts for desktop storage needs.
I conducted two separate test phases, one using MCS and one using PVS, provisioning 2,000 persistent VDI desktops in both cases. In each phase, I collected I/O metrics during the desktop provisioning process and in a boot storm scenario in which all 2,000 seats were booted simultaneously.
Provisioning and boot storm results
As shown in the table below, the NetApp FAS8060 system greatly accelerated MCS provisioning and boot times. MCS provisioning took less than 41 minutes to provision all 2,000 desktops and only 16 minutes to boot the desktops in a boot storm scenario. PVS took 88 minutes for provisioning and 25 minutes to boot. Since PVS must also copy write cache disks during the provisioning process, it makes sense that provisioning takes longer. As expected PVS featured substantially lower IOPS, averaging 227 versus 41K for MCS during the initial provisioning process and 11K versus 68K during the boot storm test.
The results demonstrate that MCS is a reasonable provisioning choice for this size workload since the storage system can provide adequate IOPS. Because the NetApp All-Flash FAS8060 features such low latencies, it’s able to provide the necessary storage system performance for both MCS and PVS provisioning and boot operations.
With any desktop virtualization project, cost is always a concern, and storage can often be the most costly component in the overall solution. The NetApp FAS8060 platform and the Clustered Data ONTAP 8.3 operating system make several efficiencies possible that help to reduce storage capacity, which can in turn conserve costs.
As shown in the figure below, the NetApp storage solution enabled substantial reductions in capacity requirements for both MCS and PVS through the use of thin provisioning and always-on dedupe technology. Additional storage savings would be achieved with the additional NetApp always-on space efficiency technologies like inline zero elimination, sis clones, and inline compression. Storage capacity for the base OS decreased from 40GB to 24GB for both MCS and PVS. Without thin provisioning and deduplication applied, MCS would require about 2.8TB to support differencing disks for all 2,000 desktops. With thin provisioning and deduplication only about 1.2TB is needed, a reduction of 60%. Because PVS provisioning requires storage for write cache disks, 10TB would have been needed without the use of these efficiency features. With thin provisioning and deduplication, PVS requires only 1.2TB—an 88% savings.
This testing exercise demonstrates how the NetApp All-Flash FAS8060 system can supply low latencies and satisfy IOPS demands for either MCS or PVS provisioning of 2,000 persistent VDI desktops. Whether MCS or PVS is used, NetApp always-on storage efficiency technologies reduces storage requirements. (A new feature in Data ONTAP 8.3, Advanced Drive Partitioning, also increases capacity for data since it eliminates the need for a dedicated root aggregate.)
Citrix and NetApp engineers collaborate on many reference architectures and test scenarios to help customers design responsive, highly available, and cost-effective desktop virtualization solutions. Many of these solutions are documented—including the white paper about this testing exercise—to provide best practices and guidelines to help you build a successful solution architecture. As always, it’s recommended that you deploy a proof-of-concept to collect metrics with typical workloads to help you appropriately size a deployment.
To read more about the NetApp All-Flash FAS8060 system and the results of the XenDesktop testing, see the full test report white paper and also NetApp’s recently published Reference Architecture TR-4243.
— Rob Briggs, Principal Solutions Architect with Citrix Worldwide Alliances