Parallels VDI Solution White Paper VDI SIZING A Competitive Comparison of VDI Solution Sizing between Parallels VDI versus VMware VDI www.parallels.com Parallels VDI Sizing. 29
Table of Contents Overview... 3 Key findings... 4 Parallels VDI Solution Overview... 5 What is Parallels VDI Solution?... 5 Testing Environment... 7 Methodology... 8 Results for the Light Worker Workload... 1 Results for heavy worker workload... 12 Virtuozzo 4. VDI sizing. 28 Parallels 2
Overview IT organizations find incredible value in benchmarks and the ability to provide side by side comparisons of tests and performance characteristics of different VDI technology, architectures and solutions, using the exact same methodology and criteria. With that goal in mind, this paper attempts to replicate as accurately and fairly as possible the essence of the VMware VDI Sizing white paper 1. This paper outlines the tests, configurations and how the final results were obtained. Four Components of a VDI Solution Connection Broker Management Tools Parallels Virtual Desktops Remote Users Firewall Corporate Users Management Tools This is a diagram summarizing the Parallels VDI Solution configuration. For more information on the Parallels VDI Solution, technology and benefits, visit the Parallels VDI Solution website at www.parallels.com/solutions/vdi/. 1 http://www.vmware.com/pdf/vdi_sizing_vi3.pdf Virtual Desktop Infrastructure 3
Summary of Key findings The VMware VDI white paper had some very specific use cases, a heavy and light workload. For the same workloads and testing criteria that were replicated and reproduced for this white paper, the following are some of the findings: Parallels VDI solution can support up to 14 desktop images per server under the «Light Worker Load». In comparison, the VMware VDI solution tests show a result of just 42 desktop images per server 2. Because the VMware tests were performed on older hardware, it is possible to assume the VMware VDI solution could attain more desktops per server, a generous estimation could be as many as 6 desktop images per server, an increase of more than 4% in server density. Parallels VDI solution can support up to 13 desktops 3 under the «Heavy Worker Load». The VMware hypervisor tests show just 26 desktops per server1 but again should be slightly more when taking newer hardware into consideration, a generous estimation could increase this to as many as 35 desktop images per server. Parallels VDI solution technology uses systems resource more effectively than the VMware VDI solution. Parallels VDI solution s application template technology helps to save on disk activity and memory consumption 2 http://www.vmware.com/pdf/vdi_sizing_vi3.pdf 3 We think results can be improved by improving storage type or adding more memory. It will be definitely done in next versions of the whitepaper Virtual Desktop Infrastructure 4
Parallels VDI Overview Parallels VDI solution is an integrated desktop solution providing enterprise-class control and manageability to the organization while maintaining the existing end-user experience. Parallels VDI solution is a blending of patented virtualization technology from Parallels and connection/thin client brokers to provide the following benefits: Unmatched TCO with Significant Cost Efficiencies 14 desktops per server 3x higher compression ratio than leading virtualization solutions Single software copy eliminates memory bottlenecks (on 64-bit servers) while providing a small footprint of only 5-6 Mbytes per desktop. Use of resident drivers and software provide near native server performance resulting in low operational overhead and higher consolidation ratios Greenest Technology Most energy efficient solution of any VDI solution available today Highest density of desktop images per single server hardware Ability to reuse older hardware to host desktop images and save on capital expenditures Best Manageability Parallels patented technology enables administrators to increase their span of control by managing larger numbers of desktops up to 1, desktops per server Small footprint allows creation of repeatable new desktop images via standard or custom templates with limited overhead Scheduling of automated routine maintenance procedures, such as backup, without external software or performance overhead Familiar & Improved End-User Experience Parallels patented technology provides performance benefits seen in Terminal Service implementations as well as the security achieved in traditional hypervisor deployments Ability to provide custom desktops through multiple levels of user access for resources and applications Ability to create management and provisioning templates by group Provision applications to specific user requirements Ability to change system resources including hard drive, CPU, and memory real-time to meet changing processing requirements. Before going into details on the test results and comparisons, it is important to understand the architectural and technology differences. The technology muscle of Parallels VDI is the leading operating system virtualization solution, Parallels Virtuozzo Containers R, for Windows and Linux ecosystems. Within the typical software stack, Parallels virtualizes at the operating system, or kernel, layer presenting a single operating system on the physical server. Parallels container virtualization isolates the workload and dependent services liberating the operating system kernel and resource management subsystem. Virtual Desktop Infrastructure 5
The result is centralized control of computing resources that are allocated on demand ensuring workloads have access to all available resources. Parallel s container architecture eliminates under utilization or unavailable system resources, which occurs with virtual machine solutions such as VMWare. Other benefits of container virtualization include: Namespace Isolation each domain has its own namespace, objects within domain could not discover and access objects in another domain Data Isolation data, created within one domain should not be visible or modifiable by other domains Resource Isolation each domain gets its own share of system resources in controlled fashion. Performance Isolation one domain should not be able to influence application performance in other domains Fault Isolation each domain has its own namespace, objects within domain could not discover and access objects in another domain Container virtualization is ideally suited for high I/O intensive applications such as VDI. Container virtualization s single operating system instance is preferred for environments where lockdown desktop images are often limited to a single controlled version. Container virtualization places more granular isolation boundaries around the workload and services freeing up the resource management subsystem for all containers. The result is centralized control and allocation of workloads on demand ensuring that all available resources are used nothing wasted. Virtual Desktop Infrastructure 6
Testing Environment Hardware: Hosting Server for Virtual desktops: HP DL38/G5 2 CPU x Quad code Xeon 2. GHz 16 GB of RAM 2x75 GB SCSI drives in RAID, Smart Array P4 SCSI controller, 2 x 1 GBit HP NC373i adapters Clients (2): HP DL38/G5 2 CPU x Quad Core Xeon 2. GHz 4 GB of RAM Software: OS for the server and clients: Microsoft Windows 23 SP2 R2, Enterprise x64 Edition Page file size was set to 4 4 Gbytes Parallels Virtuozzo Containers 4. for Windows without tools. Each container has 1 vcpu. MS Office installed as template Microsoft Office 23 (as application template for the containers) Domain controller, IIS, second Load Generator Network Switch Primary Load Generator Virtuozzo Server running desktops Virtual Desktop Infrastructure 7
Methodology We used Terminal Server scripts and tbscript.exe utility 4 to generate a load against the tested server. Each client (Terminal User, running on Load Generator server) connects to corresponding desktop (a Parallels VDI Solution container based on the Parallels Virtuozzo Containers technology, running on the tested server) and performs some typical actions. Of course we can t fully replicate a server s behavior during a business day, but we can estimate the approximate number of containers with desktop images that can run under some typical workload. Real numbers can be different due to many factors, such as sporadic user activity, applications type, hardware configuration and so on. According to general guidelines, we choose two workload types: Light worker workload Description: Workers that input data into computer systems, for example: transcription, typing, order entry, clerical work, and manufacturing. Details: Connects to the desktop and logs in Performs the following actions in a loop: Start Internet Explorer. Load a page with heavy graphics. Close Internet Explorer. Start Word. Type a small document. Close Word. Start Excel. Open an Excel sheet. Close Excel. Start Internet Explorer. Load a page with heavy graphics. Close Internet Explorer. Heavy worker workload Description: Workers who gather, add value to, and communicate information in a decision support process. They enter data at a moderate speed. Example job tasks include marketing, project management, sales, desktop publishing, decision support, data mining, financial analysis, executive and supervisory management, design, and authoring. Details: Connects to the desktop and logs in Performs the following actions in a loop: Start PowerPoint. Load a massive presentation, browse the slides. Close PowerPoint Start Internet Explorer. Browse three different web pages. Close Internet Explorer Start Command Prompt. Do a directory listing Start PowerPoint. Load a massive presentation and browse the slides. Close PowerPoint Start Excel. Open an Excel sheet. Close Excel Start Word. Type a small document. Close Word 4 The tbscript.exe and script examples are available as a part of Windows 23 resource kit Virtual Desktop Infrastructure 8
System parameters and termination point: To determine maximum number of containers that can run under pre-defined load we performed the following: Start an OS container, wait for two minutes, start corresponding client to simulate load, wait for three (for less than 1 containers) or five (for more than 1 containers) minutes, start next container During the p.1 we monitored the basic system parameters CPU usage, committed memory, disk and network activity, number of context switches, page file usage and some others. Start Command Prompt. Do a directory listingcontainer We continuously measured the time of a single load cycle inside a canary container. The canary container is where the test s execution times were measured System parameters and termination point: Client operations terminated by timeout We observed a huge (>1%) increasing of load cycle time in the canary container System behavior becomes unstable and/or system parameters are in critical area for example, very high CPU load (8% 9%) or saturated disk/network Virtual Desktop Infrastructure 9
Results for the Light Worker Workload According to the defined methodology we monitor basic system parameters to ensure that the system is stable, working and can support more users. CPU Usage %% 1, 9, 8, 7, 6, 5, 4, 3, 2, 1,, 2 4 6 8 1 12 14 16 Number of running Desktops Termination point We found that the system s behavior become unstable and clients have problems with load generation (many clients become idle for some time due to busy server actually CPU bursts up to 1%) when average CPU load is more than ~8%. As one can see from the above graph, in this case termination point is about 14 Desktops (OS containers). Commited memory Gbytes 25, 9, 2, 7, 15, 5, 1, 3, 5, 1,, 2 4 6 8 1 12 14 16 Number of Desktops Memory consumption is almost linear. It is important to underline that for 14 Desktops support we need just ~16 Mb per Desktop. Note, that it is total memory size needed for Desktop support, in case of Hypervisors (like VMware or Xen) you need about 256-384 MB per Windows 23 VM plus additional memory for virtual machine monitor to support each running VM. Virtual Desktop Infrastructure 1
Disk total IO Mb/sec 25, 9, 8, 7, 6, 5, 4, 3, 2, 1,, 2 4 6 8 1 12 14 16 Number of running Desktops Disk speed may be a possible bottleneck with the heavy worker thread. The termination point was reached right after burst disk activity. It s possible that with faster disks, the light worker workload results could be hit for this test as well. Virtual Desktop Infrastructure 11
Results for heavy worker workload AAccording to the defined methodology we monitored basic system parameters to ensure that the system is stable, working and can support more users. Additional control was performed for clients we periodically monitored their activity and checked whether they could connect to the system and generate load. CPU Usage %% 1 9 8 7 6 5 4 3 2 1 2 4 6 8 1 12 Number of running Desktops Termination point 14 Despite low CPU usage, the system had problems with client connections and load generation at about 13 desktops. It doesn t mean that recommended CPU load for VDI server should be less than 6%, but it means that most probably there is some bottleneck like storage or memory. Commited memory Gbytes 25, 9, 2, 7, 15, 5, 1, 3, 5, 1,, 2 4 6 8 1 12 14 Number of Desktops Memory consumption is typical almost linear. All the statements about memory using benefits also apply to the light worker workload. Virtual Desktop Infrastructure 12
Disk total IO Mb/sec 25, 9, 8, 7, 6, 5, 4, 3, 2, 1,, Number of running Desktops Termination point 2 4 6 8 1 12 14 Disk speed may be a possible bottleneck with the heavy worker thread. The termination point was reached right after burst disk activity. It s possible that with faster disks, the light worker workload results could be hit for this test as well. Virtual Desktop Infrastructure 13