PvD Image Update Performance Guidance and Recommendations This article contains guidance and recommendations for PvD Image Update Performance.
Contents Introduction... 2 What an image update operation includes... 3 Connection settings that affect broker throttling... 3 Parameters measured in the tests... 3 Base image snapshots used in the tests... 4 Detailed test results for 7.6 and 7.6.1 releases... 4 Performance Improvement achieved in 7.6.1... 5 General considerations and guidance for optimizing power actions... 5 Appendix A - Test environment description... 7 Page 1 of 7
Introduction This article describes the performance testing in a XenDesktop 7.6.x deployment of an image update operation on a machine catalog containing desktop OS virtual machines (VMs) with Virtual Delivery Agents (VDAs) that have Personal vdisk (PvD) enabled. This document has been updated with the test results of PvD version 7.6.1. The testing on 7.6 was focused on determining the optimal power action values relevant to an image update operation, and benchmarking the performance. The PvD 7.6.1 release contained performance improvements specific to an image update operation. This improvement has been validated on identical hardware and software against which the 7.6 measurements were conducted. Test result summary Using a five-node XenServer hypervisor cluster with shared NetApp storage for the PvD disks, and a machine catalog containing 200 VMs with PvD enabled on their VDAs: When using Machine Creation Services (MCS), a full image update took 1 hour and 44 minutes to complete using PvD 7.6 and 1 hour and 16 minutes using PvD 7.6.1. General Guidance for PvD Throttling Settings The following is a rough rule of thumb for defining settings that affect image update in end-user deployments so that the system is not bottle-necked by IOPs or very under-utilized, and completes in a reasonable time frame. Note that these are not necessarily optimal settings, but they will provide decent performance. Determine the total IOPs supported by the storage. This should be manufacture-specified or calculated. Using 350 IOPs per VM, determine the number of VMs that should be active at any point in time on the storage. Do this by dividing total IOPs by 350. Use this number of VMs calculated above for the Absolute Value in the Simultaneous Personal vdisk inventory updates setting. Set the percentage value to a 100%. The lower of the two values will take effect. Example: If the total IOPs supported by end-user storage is determined to be 14000 IOPs, then the setting for the number of VMs will be 14000 IOPS / 350 (IOPS per VM) = 40. Leave all other throttling settings at their default. See Detailed test results for result details. See Test environment description for infrastructure and test details. Page 2 of 7
What an image update operation includes After you update a master image that is used in a machine catalog, the image update process upgrades all VMs in that machine catalog to that new master image. If you are using Machine Creation Services (MCS), the process begins with taking a PvD inventory. If you use Provisioning services, preparation is a manual operation. Then, for each machine in the catalog: 1. The machine is powered on/restarted and placed in a preparing state. (Powered-off machines are also placed into the preparing state, but the PvD image update will not begin until the machine is powered on.) 2. The PvD is updated with the new master image. 3. The machine is restarted. 4. The VDA registers with the Controller and is then ready for use. Connection settings that affect broker throttling The Citrix Broker Service controls the power state of the machines that provide desktops and applications. The Broker Service can control several hypervisors through a Delivery Controller. Broker power actions control the interaction between a Controller and the hypervisor. To avoid overloading the hypervisor, actions that change a machine s power state are assigned a priority and sent to the hypervisor using a throttling mechanism. The following settings affect the throttling. You specify these values in Studio by editing a connection (Advanced page). Simultaneous actions (all types) - The maximum number of simultaneous in-progress power actions allowed. This setting is specified as both an absolute value and as a percentage of the connection to the hypervisor. The lower of the two values is used. (Default = 100 absolute, 20%) Simultaneous Personal vdisk inventory updates - The maximum number of simultaneous Personal vdisk power actions allowed. This setting is specified as both an absolute value and a percentage of the connection. The lower of the two values is used. (Default = 50 absolute, 25%) Maximum new actions per minute - The maximum number of new power actions that can be sent to the hypervisor per minute. Specified as an absolute value. (Default = 10) Parameters measured in the tests Total response time the total time it takes to complete the image update: the time difference between the start of the image update (called T1) and when the VDA on the last machine in the catalog registered with the Controller (called T2). The total response time = T2 T1. The time each VDA registered is available in the application event logs of that machine under the Citrix Desktop Service provider with event ID 1012. Storage average IOPS Average of the total IOPS recorded in the test environment using a polling interval of seven seconds. Page 3 of 7
Storage maximum IOPS Maximum of the total IOPS recorded in the test environment using a polling interval of seven seconds. Base image snapshots used in the tests Two snapshots were used: the first snapshot was used to create the initial machine catalog, and the second snapshot was used for the image update operation. Common to the initial and updated images: o Operating system: Windows 7 SP1 x86 o Memory: 1.5 GB o Storage primary disk (C drive): 24 GB Applications o The initial base image contained the applications: Adobe X, Firefox 10, and Notepad++ 6.4.1. o The updated base image contained the following applications (these applications comprised a large number of files and registry entry additions to the base VM, resulting in a larger image update operation): Visual Studio 2012, SQL Server 2012, and Microsoft Office 2013 Professional. Detailed test results for 7.6 and 7.6.1 releases The optimum power actions settings for the test environment used the default values for all settings except for the absolute number of simultaneous Personal vdisk inventory updates, which was 60 (default = 50). The following table compares the values obtained when using MCS for PvD 7.6 and 7.6.1 Characteristics MCS Data on 7.6 MCS Data on 7.6.1 Total time 1 hour 44 minutes 1 hour 16 minutes Average storage IOPS 14987 13725 Maximum storage IOPS 25035 25958 Test Results for PvD version 7.6 using Provisioning Services Characteristics PVS Data on 7.6 Total time 1 hour 39 minutes Average storage IOPS 9397 Maximum storage IOPS 20977 The following graph illustrates the typical distribution of Total IOPS (orange line) and the number of machines being updated simultaneously (blue line) for the duration of the image update operation. Although the time interval for which IOPS measurements were made is in seven second intervals, the graph is plotted in three minute markers for readability. Page 4 of 7
Total IOPS 25000 20000 15000 10000 5000 0 12:59 PM 1:02 PM 1:05 PM 1:08 PM 1:11 PM 1:14 PM 1:17 PM 1:20 PM 1:23 PM 1:26 PM 1:29 PM 1:32 PM 1:35 PM 1:38 PM 1:41 PM 1:44 PM 1:47 PM 1:50 PM 1:53 PM 1:56 PM 1:59 PM 2:02 PM 2:05 PM 2:08 PM 2:11 PM 2:14 PM 2:17 PM 2:20 PM 2:23 PM 2:26 PM 2:29 PM 2:32 PM 70 60 50 40 30 20 10 0 # of Desktops Total IOPS Image Update In Progress For 7.6, in addition to testing the catalog containing 200 VMs, tests were run on other catalogs of different sizes to measure the Total Response Time. These results are shown in the following graph. Total Response Time Distribution in 7.6 450 400 350 300 250 200 150 100 50 0 1 2 3 Number of Desktops 100 200 400 Total Response Time(minutes) 50 104 175 Number of Desktops Total Response Time(minutes) Performance Improvement achieved in 7.6.1 With the latest software optimizations in place in version 7.6.1 we were able to achieve a performance improvement of ~27% over 7.6 for the total response time. We were also able to achieve a higher absolute personal vdisk inventory updates value for the PvD preparing state. For 7.6 the optimal value was calculated to be 60. In 7.6.1 the optimal value has been calculated as 80 which indicates that a higher number of desktops are able to update simultaneously using the same resources General considerations and guidance for optimizing power actions The supported hypervisors (XenServer, VMware, and Hyper-V) may have differing capabilities of handling power actions. To determine your optimal connection throttling settings, you need to know the Page 5 of 7
Input/Output operations per second (IOPS) based on the capacity of the hypervisor s storage infrastructure. The higher the IOPS handling capability of the storage, the better the performance. To identify optimal connection throttling values: 1. Using the default values, measure the Total Response Time for an image update of a test catalog. Also measure the IOPS of the hypervisor storage during the image update. This data can serve as a benchmark for optimization. The default values may be the best setting; alternatively, the system might max out of IOPS, which will require lowering the setting values. 2. Change the Simultaneous Personal vdisk inventory updates value as described in the next step, keeping all other settings unchanged. 3. Increase the value by 10 and measure the Total Response Time after each change. Continue to increase the value by 10 and test the result, until deterioration or no change in the Total Response Time occurs. If no change in the PvD power action value occurs, decrease the value in increments of 10 and measure the Total Response Time after each decrease. Repeat this process until the Total Response Time remains unchanged or does not improve further. This is likely the optimal PvD power action value. In the test environment, the best performance was achieved with the PvD power action set to 60. All other values for this setting resulted in increased Total Response Time. 4. After obtaining the optimal PvD power action setting value, tweak the values for Simultaneous actions (all types) and Maximum new actions per minute, one at a time. Follow the procedure described in the previous step to test different values. Page 6 of 7
Appendix A - Test environment description The following diagram shows the infrastructure used for testing. The following table lists the VDI hardware and configuration. Servers XenDesktop Controller VMs containing VDAs Type: HP SL390s Gen 7 - XenServer pool of five nodes supporting 200 VDAs Processor: 2 x vcpu RAM: 6 GB Processor: 2 x vcpu RAM: 1.5 GB Processor: 2 x Intel Xeon X5650 CPU @ 2.67 GHz (6 cores) RAM: 96 GB Hypervisor: XenServer 6.2 with updates to XS62ESP1008 installed. Platform: Windows 2008 R2 Platform: Windows 7 x86 SP1 Storage configuration: NetApp FAS3240 used as the shared storage repository for the VDA and PvD. Total usable size: 4.5 TB Two volumes configured: one for storing the personal vdisk and the other for VDA disks Connection to the XenServer pool through an Extreme 10 Gbps switch that isolates the management and storage traffic All tests were completed in a XenDesktop 7.6 deployment on a machine catalog containing 200 machines, each with a VDA for Desktop OS installed and PvD enabled. The tests used the following provisioning tools using XenDesktop 7.6; at least three runs were executed for each provisioning tool type, and averages taken for the final reported data: Machine Creation Services (MCS) Base images are maintained through snapshots. Provisioning Services (PVS) Provisioning Server versioning maintains the image update base images. Page 7 of 7