X-Pod for VDI. Validated Reference Architecture

Transcription

1 X-Pod for VDI Providing a simple yet best of breed converged infrastructure for virtualized desktop solutions Enabled by: X-IO ISE 740 Hybrid Storage Array VMware Horizon View 5.3 Cisco UCS Validated Reference Architecture August 2014

2 Table of Contents Introduction... 3 Key takeaway... 3 Executive Overview... 3 Why VDI?... 4 VDI Business Benefits... 5 Flexible Desktop Environment... 5 Better, Easier Desktop Management... 5 Desktop by Template... 6 Security and Compliance... 6 BYOD Support... 6 Virtual Desktop Implementation Risks... 7 Performance & Capacity: Not a Trade-Off... 7 Reliability and Redundancy... 7 Deploying and Expanding Desktops Pools... 8 Redefining Steady-State Operations... 8 VDI Planning Questions to Ask... 9 Solution Overview Components Solution Architecture Hardware Components Software Components Cisco Unified Compute System (UCS) VMware vsphere VMware Horizon View 5 Environment Architecture VMware vcenter Operations Manager X-IO ISE 740 Hybrid Storage Array Performance Analysis of Tested configurations Test Methodology Workload Analysis Virtual Desktop Deploy Operations Virtual Desktop Boot Storm Conclusion Contact X-IO technologies Appendix Appendix A: Detailed Analysis of vsphere Performance Graphs, Login VSI 500 User Run Appendix B: EXAMPLE Login VSI Target Desktop Pool Configuration X-Pod for Horizon View 5.3 2

3 Introduction This white paper provides a primer for the X-Pod for VDI converged infrastructure solution, powered by X- IO storage. X-Pod is a reference architecture designed to deliver a repeatable, high-performance virtual desktop infrastructure. It utilizes an industry-leading VMware Horizon View environment, hosted on Cisco Unified Computing System (UCS) Blade Servers, with storage housed on a single, highperformance, X-IO ISE 740 Hybrid Storage Array. The environment described herein has been extensively tested by X-IO and verified by VMware. This document is intended to provide insight into the components, the architecture, and the performance required to meet the demands of a 500, 1000, and 1500 seat virtual desktop infrastructure (VDI) deployment. Key takeaway This paper details the extensive testing that X-IO has performed and provides guidance as to the performance requirements of various operations based on the X-Pod core components of Cisco UCS server and networking hardware together with the X-IO Intelligent Storage Element (ISE) systems. This validated testing demonstrates that this solution is capable of delivering a high-performance desktop experience at up to 97% concurrency. It is designed to act as both a primer for VDI deployments and as a technical overview of the X-Pod architecture. Those readers who are comfortable with the business and technical benefits of deploying VDI together with the pitfalls of implementing such an infrastructure should progress to the Solution Overview section of this document. Executive Overview When discussing VDI architectures with customers and solution providers, X-IO have noticed a common trend of two concerns. Firstly whether the technical statistics and benchmarks provide are real-world or marketing-hyped numbers. Secondly whether VDI architects should be utilizing hyper-converged systems or using best of breed components. It is therefore worth briefly covering these points in turn. When it comes to VDI design, there are many pitfalls that can occur. This has led to a frightening statistic that for every successful VDI implementation, there are 7 to 10 that fail, the vast majority of these being due to an inappropriate storage design. X-IO have therefore worked with real-world benchmark tools such as Login VSI to ensure that any benchmarks carried out are true reflections of user actions rather than synthetic workloads applied by I/O testing tools. All of the user counts reflected in this document are actually high-utilization loads and if anything, are worst-case numbers rather than optimistic based upon assumptions or unrealistic scenarios (e.g. zero user data or ridiculously high data de-duplication forecasts). With regards to the argument for hyper-converged systems, the systems out there today undoubtedly have their place, however they need to be precisely architected for today s workload and are relatively inflexible for future growth or change of use. This has often led to a difficult decision for architects as to X-Pod for Horizon View 5.3 3

4 whether to endure the complexity and risk of using best of breed components or to compromise and deploy a pre-converged, collapsed stack. With the arrival of converged solutions such as X-Pod, best of breed components can be deployed, however using pre-architected and tested blueprints that give the peace of mind previously only found with hyper-converged systems. Why VDI? Virtualization of servers and IT infrastructure has been well established in the business landscape. Operating expense (OPEX) cost reductions are routine in a virtualized data center due to the reduced number of physical servers, more centralized management tools, energy savings, and many other factors. This cost savings makes desktop virtualization such a promising opportunity to transform a major cost for IT organizations. Unfortunately, virtualized desktop workloads are unlike most of the server virtualization workloads with which organizations have experience. While steady-state desktop operations are what is commonly sized for (for example, IOPS/desktop), it is the non-standard operations and misconfigured sizing operations that eventually cause poor user experience and make VDI such a challenging, and expensive, solution to design and implement. By leveraging desktop virtualization solutions such as VMware Horizon View, IT organizations can provide their customers with a superior desktop experience while decreasing management cost and increasing flexibility of the organization. IT organizations can now increase the leverage of their IT staff by consolidating desktop management and hardware into the enterprise virtualization computing model. For example, relatively few IT staff can manage hundreds (if not thousands) of desktops for patch management or application upgrades. Using tools such as VMware Horizon (with View) to manage snapshots, gold images can be quickly reverted to a known state if required, which reduces capacity as only data changes are stored for each snapshot. By contrast, traditional methods of patch management require excessive involvement from the IT organization across hundreds (if not thousands) of physical machines that could be spread around the world. In addition to desktop virtualization, tools are now available to virtualize applications as well. End users can be provisioned individual applications that can be delivered to any device, whether that is a laptop, tablet, or smartphone. As the bring your own device (BYOD) phenomenon continues to grow, users demand more flexible access to the resources they need to perform their jobs. Policies that prohibit employees from using personal devices, and that the company does not pay for, are becoming more difficult to defend. Implementing a virtual desktop solution can be immensely beneficial to the organization, but if implemented incorrectly, it can be an equally impressive failure. User experience is the gold standard that any solution will be held to, because users will evaluate the new solution based on the physical desktop that was just replaced. Proof of Concept (POC) testing is essential when evaluating any new solution, because the process of configuring the test gives essential information about how the system will perform. Tools such as Login Virtual Session Indexer (Login VSI ) are critical for simulating like production workloads and are a worthwhile investment as part of evaluating different solutions. Inadequately designed VDI implementations often perform well in the POC phase but are overwhelmed when placed into production. In a majority of these cases, storage performance is often the cause of poor user experience, is the most often undersized for performance, and can be the most expensive single component of any VDI solution. X-Pod for Horizon View 5.3 4

5 VDI Business Benefits Properly sized and implemented, a VDI project can provide users with a whole desktop experience that surpasses their physical desktop machine. However, the benefits to the business are similar to server virtualization but are applied at an order of magnitude greater scale. Enterprise-class computing hardware can be applied to end-user computing (reliability/availability/performance) and can leverage large-scale virtualization management techniques (capacity/scale). This can allow for greater leverage of existing IT staff, enable higher levels of reliability (including disaster recovery), and reduce overall ongoing costs to the business. While VDI can be a tremendous advantage for the business, data storage costs are the single largest cost component of the solution and the most common cause of poor user experience and failed implementations. Listed below are some advantages to the business and the role that storage has to play in each. Flexible Desktop Environment Many types of desktop needs are seen across various organizations and industries. Based upon the load pattern and use case, different kinds of desktop virtualization options are available. In fact, it is likely that multiple types of implementations exist in a single, larger organization. For example, the resource requirements for graphic designers, software developers, Microsoft Office power users, and executives are quite different than the requirements for a call center or kiosk desktop. The power users will generally have higher CPU and memory requirements and will likely require that the desktop be persistent. Power users will expect any changes to their desktop to be preserved and in the same state they left it when logging off. In a call center (or kiosk) configuration, users expect to log in to any desktop and have the exact same experience with nothing preserved on the desktop. If one desktop has an issue, the user simply moves to another desktop instance and tries again. There is usually a standard suite of software that these virtual desktops utilize to support the business functions. This is especially true if, for example, cloud-based customer resource management (CRM) service or Office365 are being used. Better, Easier Desktop Management With physical desktops, the IT organization has to go out and physically touch each desktop in some cases to remediate a problem. This geographic dispersion, even if within the same office building, increases the staffing requirements to manage the end-user desktops and increases the time required for break/fix functions. There is also an increased risk of data theft when users store corporate data on physical desktops that reside anywhere in the organization. Desktop virtualization allows for centralized management of the most commonly touched component in the desktop solution, the desktop itself. Backups can also be more effective, since all user data is kept in the datacenter even though it looks to the user like an attached, physical disk housed on local storage. This eliminates many of the issues seen with trying to back up hundreds and thousands of remote, physical desktop machines. Network interruptions, issues with individual operating system bugs, failing physical hard drives, and so on are all mitigated by utilizing a common storage platform. In addition, since the effective storage for each user desktop is deduplicated by the VDI layer (VMware View Composer Linked Clones), the amount of storage actually needing to be backed up is vastly reduced and quickly recovered. VMware has for years provided a rich API to enable storage partners to better integrate with the VMware ecosystem. This enables the virtualization administrator (vadmin) to quickly and efficiently assign storage X-Pod for Horizon View 5.3 5

6 resources where required while the complex provisioning of storage resources up through the different virtualization and hardware layers is done by software. This greatly leverages the amount of storage the vadmin can effectively manage and ensures that best practices are followed for storage configuration, thereby reducing risk. Desktop by Template IT organizations can create a library of virtual desktop templates, each of which can be carefully tuned and configured for the business need. These templates can then be used to provision virtual desktop machines very, very quickly in minutes in most cases. For example, if an organization routinely uses contractors for a few different types of functions, a gold image template can be created for each, which can include the appropriate operating system, security settings, and all the applications contractors would need to accomplish their tasks. When a new contract employee is added, the IT staff need only deploy a new virtual desktop based upon the appropriate template. This can be done literally in minutes, but this can also generate an enormous amount of storage traffic or I/Os per second (IOPS) with a relatively small number of deploy operations. NOTE: Deploying a desktop is one of the most performance-demanding operations that storage will encounter in a desktop virtualization solution. Performing these operations during steady-state operations can put an abnormally large strain on the storage infrastructure and can impact the user experience of all other users on the system. Security and Compliance VDI allows an IT organization to have much better control over corporate data. As such, it makes the job of implementing consistent, common security and compliance features easier. Depending upon the industry the organization is in, compliance and security concerns can be important or strictly mandated. With physical desktops and mobile devices, corporations face the increased risk of data loss and theft of any device where data is stored locally. Desktop virtualization enables employees to work with data securely on centralized corporate resources. In some cases, users can work from any device in any location with their data meeting the security requirements of the organization. Depending on the role of the desktop users, certain data regulations may apply that require data separation and isolation (e.g., financial, legal, and medical). This requires separate storage devices that must be able to support high IOPS and capacity levels that the separated pools require. Modular-based storage systems are inherently designed to accommodate this data security requirement. BYOD Support Increasingly, organizations are supporting (and are being demanded to support) bring your own device (BYOD) functionality. The workforce is changing, and the growing expectation is that applications and desktop access will be available on whatever device the employee has, from laptop to tablet to smartphone. This also has an immense benefit to the business, as the employee is providing a preferred end-point device that can be used to be productive. VMware Horizon View offers a very rich interface that allows users to access their applications and desktop on any mobile device. Individual applications can be virtualized, further enabling employee productivity and easing the demands to corporate computing resources. X-Pod for Horizon View 5.3 6

7 Virtual Desktop Implementation Risks While the list of benefits that can be realized by VDI implementation, both in manpower and in operating expenses (OPEX), is impressive, the opportunities for failure are equally concerning. This section lists several areas that can put a VDI initiative at risk. In all cases, sufficient planning for the final production load and careful monitoring of operations are required to ensure smooth operations. Performance & Capacity: Not a Trade-Off Performance has a direct effect on end-user experience, and poor storage performance sizing is usually the number one reason. Many VDI solutions are sized only for steady-state performance requirements. Common VDI operations, such as boot storms, login storms, deploy operations, recovery, and other maintenance operations, can require an enormous amount of transactional performance (IOPS) from storage systems. VDI instances have to respond with little or no discernable impact to the end user when these maintenance operations are conducted. Performance is not the only thing to consider when sizing a VDI solution, as there must be enough capacity to satisfy the requirements for user data, operating systems, application data, and user persona. While there are several techniques to increase the amount of effective capacity a solution can provide (VMware View Composer Linked-Clones), there is still an underlying requirement for capacity performance. Unlike regular file server class capacity, this data has performance requirements that go across all of the data (e.g., there is little dead data ). This broad requirement means that while there may be areas of high-performance concentration (base images or replicas), the rest of the data has a high IOPS/TB requirement as well. This rest of the data is what the individual VDI instances have to work with, so any slowdown in capacity will directly affect the end-user experience. Solutions that rely on calculated deduplication and compression techniques will see degrading performance ratios as the capacity is consumed, as there are limited amounts of processing power and memory capacity in the storage controllers (i.e., the number of calculations increases with increasing capacity). According to VDI support organizations, storage systems account for 80 90% of the VDI performance issues reported. Often this is due to read and write latency in the storage system. Reliability and Redundancy No virtualized datacenter is sustainable without reliability and redundancy. In the past, services running in the data center did not necessarily represent a disaster if there was a service disruption. For example, a service event to the CRM applications, or services, would be painful to the organization if they were sluggish, non-responsive, or unavailable for a period of time. Users could still use their desktop for other productive functions, such as Microsoft Office applications or researching on the internet. Virtual Desktop Infrastructures, on the other hand, can represent a true disaster if they become sluggish or non-responsive. The user is no longer impacted by only one or two applications but by the very desktop itself. Imagine the majority of the workforce reduced to typing 10 or 20 words per minute or login times in excess of 30 minutes. Any slowdown in the VDI ecosystem has a very public result. Storage systems incur a tremendous performance penalty when performing disk drive recovery operations, and these have a dramatically negative impact to the end-user experience. Even if data is separated into groups or sets, centralized storage controllers are now responsible for managing recovery operations in addition to ongoing operations. Storage companies would not generally consider this a single point of failure, but users will. X-Pod for Horizon View 5.3 7

8 Deploying and Expanding Desktops Pools Over the course of the reference architecture testing, various user configurations were tested. At one point in the testing, three desktop pools of 500 users each were expanded by 10 desktops each. This was done while all of the desktops were powered on, but no workload was being applied (idle). Below is the graph of the IOPS from the single ISE 740 system that was tested. Regular spikes in IOPS can be seen to reach over 12,000 IOPS, and the entire operation lasted for roughly 10 minutes. Performing this activity during the production period would have a significant impact to other functions that required storage performance, unless the storage system had enough headroom to accommodate the additional load. Traditional storage solutions that can accommodate this level of activity are cost prohibitive and are often undersized for this level of performance. Figure 1 - Deployment Operations and High IOPS The majority of the deployment time and performance requirements came from the Sysprep operations for the Windows 7, 32-bit operating system. Cloning of the base images and creation of the View Composer linked clones represented ~1/3 of the deployment time for the 30 desktops mentioned above. In testing with large numbers of desktops (500), the time and performance required for the Sysprep functions (not cloning the base image) comprised the vast majority of the deploy operation time and performance required. Redefining Steady-State Operations When sizing a VDI environment, many different operations must be planned for. Accommodating hosted desktop users during steady state with a low-latency experience has to be satisfied during maintenance operations. Boot storms and login/logout storms of relatively few numbers of users can push a storage system designed only for steady state well past its breaking point. These operations are part of the normal desktop access activity for users, and as such they should be included in any storage sizing discussion for normal operations. In testing with Login VSI, storage performance (IOPS) required during the user login phase was two times the steady-state workload. Considerable thought should be given to user behavior patterns and how many users will be concurrently initiating connections to the desktop environment. Login VSI is capable of adjusting the rate at which users log in to the environment for the test and is an excellent method for exploring the performance of the solution during various login activity levels. In the testing X-IO performed, user login rates of 16/min, 25/min, and 33/min were tested. The single greatest impact to the X-Pod for Horizon View 5.3 8

9 login times was the high compute (CPU and RAM) utilization of the UCS blade servers when the majority of the users had connected to the solution. VDI Planning Questions to Ask Proper planning is essential to a successful VDI roll-out. Here is a non-exhaustive set of questions to help guide your investigation into a VDI pilot program and your full production program. What are the different kinds of users you will have to support (i.e., kiosks, developers, power users, tellers, knowledge workers)? What is the scope of each user type? How many simultaneous desktops will be required? What are the expected resource demands for each user (i.e., CPU, memory, disk space, network traffic)? What is the planned concurrency for your user pools and will their usage be offset from each other (i.e., shifts, geographical support)? What are the relative benefits for virtualizing desktop access for each user type? What are the relative risks for each user type, if there are access issues? What existing infrastructure can be used for the VDI implementation? For each user type, would persistent or non-persistent desktops be more appropriate? For each user type would linked clones, full clones, or dedicated virtual machines be more appropriate? What existing IT management and monitoring tools do you have in place? Will your infrastructure support iscsi? Fibre Channel? Is either preferable to you? Will the pilot program be based on actual users or will it use validation software like Login VSI? For the pilot program, what are your success criteria? How much cost will there be in extending the warranty beyond what was included with the base support period? What metrics are important to you? For example: o o o o o o Total number of IOPS Total throughput Recompose wall-clock duration Maximum latency as seen by end user Virtual machine boot time Login / logout times How will the transition from pilot program to full production implementation be done? X-Pod for Horizon View 5.3 9

10 Solution Overview This X-Pod reference architecture white paper describes a compact configuration to deliver a highperformance VDI environment that supports up to 1500 virtual desktops. In the following sections, Cisco UCS, Cisco Nexus, Cisco MDS, VMware vsphere, VMware Horizon View, and the ISE hybrid storage array are described. Components The following components were chosen for this X-Pod reference architecture. Cisco Unified Computing Systems (UCS) The underlying premise of a VDI solution is to run user desktops on powerful datacenter servers rather than on distributed physical machines. Cisco has focused on the characteristics needed to support this functionality in datacenter servers and has developed the following innovations: Extended memory Virtualization optimization, with Cisco VN-Link technology Unified I/O access and unified fabric Unified, centralized management Service profiles The bottom line is Cisco has developed and refined the Unified Computing System to specifically meet the Enterprise VDI requirements. Simplified architecture and management geared toward datacenter fulfillment of virtual desktops leads to reduced total cost of ownership (TCO) by lowering acquisition costs, lowering operating costs, and lowering ongoing operational costs. The UCS unites compute, network, storage access, and virtualization into a cohesive system. The system is integrated on a low-latency, 10-Gigabit ethernet (10GbE) unified network fabric with enterprise-class, x86-architecture servers. It is an integrated, multi-chassis platform in which all resources participate in a unified management domain. The Cisco UCS accelerates the delivery of new services simply, reliably, and securely through end-to end provisioning and migration support for both virtualized and nonvirtualized systems. For more on UCS: Cisco Nexus The Cisco Nexus 5548P is a one-rack-unit (1U), 1 GbE, 10 GbE, and FCoE access-layer switch built to provide 960 Gbps of throughput with very low latency. It has 32 fixed, 1 GbE, or 10 GbE ports that accept modules and cables meeting the Small Form-Factor Pluggable Plus (SFP+) form factor. One expansion module slot can be configured to support up to 16 additional 1 GbE and 10 GbE ports or eight Fibre Channel ports plus eight 1 GbE and 10 GbE ports. The switch has a single serial console port and a single out-of-band 10/100/1000-Mbps Ethernet management port. X-Pod for Horizon View

11 Cisco MDS Cisco MDS 9148 Multilayer Fabric Switch is a high-performance Fibre Channel switch platform. It provides low power consumption and high density with up to 48 line-rate 8 Gbps ports in one rack unit (1U). VMware Horizon View Horizon View is used to deliver virtual desktops as a service in a broad range of enterprise use cases, enabling the best user experience for maximum productivity. IT administrators can easily provision and customize the environment to comply with corporate policy and end-user needs. Desktop virtualization with Horizon View enables organizations to do more with less and adopt a user-centric, flexible approach to computing. By decoupling applications, data, and the operating system from the endpoint and by moving these components into the datacenter, where they can be centrally managed in your cloud desktop and application virtualization offers IT a more streamlined, secure way to manage users and provides agile, on-demand desktop services. X-IO Technologies Intelligent Storage Element (ISE) Consolidation and business intelligence are key themes in today s IT. Consolidation brings with it challenges in server multi-tenancy and hosted desktops while database management systems become the keys to a successful business. In both cases, fast and reliable solutions lead to a more productive and profitable enterprise. The ISE 700 series hybrid storage system keeps pace with the performance demands of today s IT without the high cost it takes traditional storage systems to keep up. The ISE 700 series provides an ideal balance of price, performance, capacity, and reliability by combining SSD and HDD into a single hybrid pool of capacity to provide SSD performance at HDD pricing. ISE outperforms systems that are up to ten times more expensive and provides an outstanding TCO by reducing operating costs associated with management, power, cooling, and datacenter footprint. ISE Manager Suite Enterprise storage is not only facing a challenging future it s facing a challenging present. Today s datacenters are increasingly heterogeneous and multifaceted. Virtualization technologies, diverse storage platforms, and cloud services create obstacles for traditional storage systems and storage management. How can you as a storage administrator be expected to efficiently and effectively manage storage in such a complex environment? The answer lies in your management tools, which must have deep integration with virtualization technologies and host operating systems while being precise, streamlined, and userfriendly. Ideal for the challenging storage scenarios of modern enterprises, ISE Manager 4.0 is the solution for today and tomorrow. It is an intuitive, flexible interface that provides simplified end-to-end storage management for multiple physical, virtual, and cloud environments from a single interface. It lets you simplify, centralize, and automate storage administration with software tailored to modern datacenters. X-Pod for Horizon View

12 Figure 2 - ISE Manager Suite Login VSI Login Virtual Session Indexer (Login VSI) is the industry standard load testing tool for virtualized desktop environments. Login VSI can be used to test the performance and scalability of VMware Horizon View, Citrix XenDesktop and XenApp, Microsoft Remote Desktop Services (Terminal Services), or any other Windows-based virtual desktop solution. Login VSI may be used to compare and validate the performance of different software and hardware solutions in an environment. Login VSI provides a method to measure the maximum capacity of an infrastructure. Simulated users work with the same applications as an average employee, such as Word, Excel, Outlook, and Internet Explorer. For more information, download a trial at X-Pod for Horizon View

13 Solution Architecture This section highlights the hardware and software configurations used to assemble this reference architecture for 500, 1000 and 1500 virtual desktops delivered with VMware Horizon View 5.3 on vsphere 5.5 U1. This environment was built on top of two Cisco UCS B-Series chassis, Cisco networking components, and X-IO 700 Series hybrid storage arrays. Figure 2 shows the logical diagram of the solution architecture for 500, 1000, and 1500 virtual desktops. The same infrastructure was used for all three tests. Figure 3 - Logical Diagram for 500, 1000, and 1500 Desktop Reference Architecture X-Pod for Horizon View

14 Hardware Components The following hardware components were leveraged to support the Login VSI test of 500, 1000, and 1500 VDI desktop loads. The 1500 desktop load leveraged the 500 and 1000 desktop clusters together that are listed in the table below. Hardware Quantity Configuration Servers Cisco UCS 5100 B-Series Chassis Desktop Cluster 2208XP Fabric I/O Extenders 2 Cisco UCS B200 M3 1 Two Intel Xeon E GHz CPU (16 cores total) 128 GB RAM Infrastructure Blade Cisco UCS B200 M3 7 Two Intel Xeon E GHz CPU (16 cores total) 128 GB RAM vsphere desktop cluster VIC Cisco UCS 5100 B-Series Chassis Desktop Cluster 2208XP Fabric I/O Extenders 2 Cisco UCS B200 M3 8 Two Intel Xeon E GHz CPU (24 cores total) 256 GB RAM vsphere desktop cluster VIC Cisco UCS 5100 B-Series Chassis 3 1 Login VSI Infrastructure 2208XP Fabric I/O Extenders 2 Cisco UCS B440 1 Two Intel Xeon E GHz CPU (24 cores total) 256 GB RAM Login VSI Server UCS-VIC-M82-8P 8 Networking Cisco Nexus Cisco MDS GB/s Fibre Channel Switch, 2 ports per ISE 700 Cisco UCS 6248 Fabric Interconnect 2 Storage X-IO ISE 710 Hybrid Storage Array 1 8Gb/s Fibre Channel for BfS Array X-IO ISE 730 Hybrid Storage Array 1 8Gb/s Fibre Channel for Login VSI Share X-IO ISE 740 Hybrid Storage Array 1 8Gb/s Fibre Channel for all Horizon View desktop pools X-Pod for Horizon View

15 Software Components See the table below for software details. Software vsphere Version ESXi 5.5 update 1 vcenter Server Operating System Microsoft.NET Microsoft SQL Server VMware Horizon View View Connection Server Operating System View Composer (installed on vcenter Server) Operating System Microsoft Software Platforms Active Directory, DNS, DHCP Windows Server 2012 Login VSI, VSIshare Server Operating System Microsoft.NET 3.5 Login VSI 4.0 Virtual Desktops: Target Desktop Operating System Microsoft Office 2010 Adobe Reader v. 11 Java DoroPDF 5.5 update 1 Windows Server 2008 R2 64-bit Standard Ed. 3.5 SP R Windows Server 2008 R2 64-bit Standard Ed Windows Server 2008 R2 64-bit Standard Ed. Windows Server 2008 R2 64-bit Standard Ed. Windows 7 32-bit SE 7 U13 VMware View Agent Virtual Desktop: Launch Desktop OS Windows 7 32 bit Cisco Unified Compute System (UCS) The Cisco UCS configuration was connected to X-IO ISE 700 storage arrays through dual, redundant paths to the MDS 9148 switches, then connected to the Cisco 6248UP fabric interconnect with dual, redundant, 10 Gbps connections. This architecture provides for a highly available and high-performance architecture. Each Cisco UCS chassis was connected to each fabric interconnect with four 10-Gbps network connections. X-Pod for Horizon View

16 The dual fabric interconnect pairs have primary and subordinate roles in this configuration. For more information about optimizing and configuring the Cisco UCS server service profiles for VDI, please refer to the following link: VMware vsphere 5 For this test configuration, the VMware vsphere 5.5 Update 1, ESXi hypervisor was deployed. All of the hosts were set up to boot from SAN (optional) using the X-IO ISE 710 hybrid storage array. Clusters The environment is organized into three main components: the VDI clusters, management cluster, and Login VSI Launcher cluster. Two VDI clusters encompass a total of 15 UCS blades and are responsible for supporting all target virtual desktops. The first VDI cluster of 7 nodes runs 500 VDI target desktops. The second VDI cluster of 8 nodes runs 1000 target VDI desktops. The two target desktop clusters together create the 1500 desktop load. 1 UCS blade (in its own cluster) is used to run all infrastructure functionality, including a Domain Controller, vcenter, VASA provider, vcenter Operations Manager, and Horizon View servers. The last cluster, the Login VSI Launcher cluster, is a UCS blade that runs 80 Login VSI Launcher desktops. VDI Clusters The two clusters under test in this reference architecture are the VDI clusters. They consist of 15 hosts, as described in the Hardware Components section above. These work together to support the test loads of 500, 1000, and 1500 virtual desktop machines. Figure 4 - VDI Cluster for 500 Desktops X-Pod for Horizon View

17 Figure 5 - VDI Cluster for 1000 Desktops Infrastructure The Infrastructure UCS blade houses all virtual machines to support the vsphere environment, the Horizon View environment, and the Login VSI testing environment. Figure 6 - Infrastructure VM for Reference Architecture Validation X-Pod for Horizon View

18 VMware Horizon View 5 Environment Architecture The VDI environment in this solution is created and managed in VMware Horizon View v To support this, the vsphere environment runs in a single datacenter with three clusters and a single vcenter instance. All hosts are running ESXi 5.5 U1. VMware View was used to construct and manage the desktop virtual machines. This test used a floating pool of non-persistent, linked clone desktop virtual machines. The details of the specific VMware View pool definition are included in the Appendix of this white paper. In addition to the linked clone method of deploying desktop virtual machines, VMware View uses VMware VSA technology. This technology allows a small amount of VMware ESXi host RAM to be used as a content-based read cache for the selected desktops read I/O operations. VMware VSA was configured at 2 GB for each VMware ESXi host and for each VMware View pool used for this test. View Pools The virtual machines in the 500 desktop VDI cluster were stored on four datastores, which are in turn stored on four ISE LUNs. Each LUN is 1 TB in total capacity, configured with RAID-1 protection and housed on a single ISE 740 hybrid storage array, as shown in the following ISE Manager and vsphere Client figures. Figure 7 - Datastores for 500 Desktop View Pools (vsphere Client) X-Pod for Horizon View

19 Figure 8 - LUNs for 500 Desktop View Pools (ISE Manager) The 1000 desktop VDI cluster is configured in the same manner. Below is a screen shot of the datastores as they are configured. Figure 9 - Datastores for 1000 Desktop View Pools (vsphere Client) X-Pod for Horizon View

20 Multiple desktop pools were configured in VMware Horizon View to accomplish the tests. The first 500 desktop test used 5 pools, each with 100 desktops, as shown in the image below. Figure 10 - Pools for 500 desktops The 1000 desktop test was done with 2 pools of 500. The 1500 desktop test was done with 3 pools of 500. Below is a screen shot of the three desktop pools of 500, used in the 1500 desktop test. Figure 11 - Pools for 1500 desktops Target VMs Base Image OS VM Hardware Version 8 vcpu 1 Memory HD Windows 7 32bit, Enterprise 1.5 GB 16 GB The target desktop virtual machines are created from a base image. This image was created as a stock Windows 7 32 bit Enterprise operating system. Windows updates were applied to bring it to the current levels. Several applications were then installed, including Microsoft Office, Adobe Acrobat Reader, and FreeMind. The view agent is installed on the gold image, and a snapshot is created. X-Pod for Horizon View

21 VMware vcenter Operations Manager vcenter Operations Manager (vcops) is a monitoring, trending, and alerting tool. It gathers metrics on all major components of the vsphere environment and automatically generates alerts if any values go outside of either manually or automatically set thresholds. Because vcenter Operations is extensible, it can learn to monitor application and product specific characteristics. This reference architecture uses two extensions to vcenter Operations: vcenter Operations Manager for View X-IO ISE Management Pack for VMware vcenter Operations vcenter Operations Manager for View The vcenter Operations Manager for View extension gathers information on the running Horizon View environment. It automatically sets thresholds for View-related metrics and also adds many dashboards to the vcenter Operations instance. X-IO ISE Management Pack for VMware vcenter Operations X-IO technologies has developed a vcenter Operations adapter extension that allows ISE storage units to participate as a full-fledged member of the vsphere ecosystem. All major performance and configuration metrics are captured and custom dashboards are added. Figure 12 - vcenter Operations with ISE Management Pack X-Pod for Horizon View

22 X-IO ISE 740 Hybrid Storage Array The ISE 740 hybrid storage array has 28.8 TB of usable capacity configured from the highest-quality, mission-critical 10K RPM SAS drives and enterprise-grade, MLC SSD into a single pool of flash-enabled storage. The ISE 740 is fully redundant with active-active controllers, each including four 8 Gb Fibre Channel ports. The ISE 700 Series includes patented Continuous Adaptive Data Placement (CADP) software, which analyzes the behavior of host I/O and automatically places hotspot data onto SSD only if measurable performance gains will be achieved. CADP runs continuously and makes data movement decisions every 5 seconds. In each of the 500, 1000, and 1500 user tests the ISE 740 delivered low-latency read and write transactions for the entire duration of the tests and project. Figure 13 - The ISE 740 Hybrid Storage Array (VMware Ready and Cisco Compatible) X-Pod for Horizon View

23 Performance Analysis of Tested configurations Test Methodology While storage vendors have for years utilized synthetic benchmark tools to simulate performance loads (Iometer, SQLIO, fio, iozone), nothing can provide more insight into performance requirements than a testing tool that performs actual end-user usage. Testing in this systems view methodology allows for many different facets of the solution to be evaluated, as different virtual desktop operations have drastically different requirements from storage. Simply using a load generator to show the performance possible from a storage array and somehow relating it to desktop virtualization workloads completely ignores the challenges that are unique to this solution design. Login VSI was used as the load generation tool, as this is capable of mimicking end-user functions, such as working with Microsoft Office applications, running Java, browsing web pages, and other common user functions. If the console is left open in one of the target desktops, this activity can be watched as the test progresses. Login VSI provides a valuable framework to gather much more information than just the main workload run, as will be detailed in the sections below. Other virtual desktop management operations were also performed as part of the setup and environment maintenance throughout the testing period. Performing these actions proved invaluable to learning about the different workloads involved in the solution. Testing to determine the scale of the ISE 740 storage array was one of the goals in the testing, and Login VSI test runs were performed with 500, 1000, and 1500 users. The medium workload setting was used for this testing series, as per Login VSI this can be considered an average workload for a virtual desktop user. Login VSI measures the end-user desktop experience and produces a metric that is a measure of the amount of desktops that a given solution could support with acceptable performance (VSImax). When VSImax is reached, that is the estimated number of desktops the solution can be expected to support. In all testing performed (500, 1000, and 1500 users), the Cisco UCS CPU utilization was the main limiting factor to achieving higher numbers of desktops. The ISE 740 was able to accommodate all of the tested user levels with no signs of a performance limit being approached. Workload Analysis Login VSI 500, 1000, and 1500 user Stead-State Workload All users in this reference configuration were logged in and simulated by Login VSI. The workload chosen for each of the remote users in all tests was medium. Login VSI produces a metric called VSImax. This is a measure of the number of concurrent virtual desktops that a given solution can support with acceptable desktop performance. Test iterations are performed, and the goal is to closely match or exceed the number of desktops that are planned to be concurrently run in production with the VSImax score. Below are the results from the 3 test iterations (500, 1000, and 1500 desktops). A clear, linear increase in the VSImax score can be seen as the number of users was increased, indicating that the VDI solution was able to accommodate the workload with no signs of encountering a performance bottleneck until 97% or greater concurrency was achieved. X-Pod for Horizon View

24 Figure 14 - VSImax for 500, 1000, and 1500 Users In the graphs above, VSImax is encountered at 492, 998, and 1466 sessions for the 500, 1000, and 1500 user tests respectively. This is due to the high resource utilization of the servers in this configuration. See Figures 15 and 16 for vcenter Operations Manager views of one of the server s CPU and memory resources during each of the 500 and 1000 user tests. X-Pod for Horizon View

25 Figure 15 - Server Resources During 500 User Test Figure 16 - Server Resources During 1000 User Test One of the things that makes the virtual desktop workload so challenging for storage solutions is the high amount of write operations that are required, especially considering the write penalties involved with RAID operations. During the steady state testing, write operations to storage were observed to be 80% of the total IOPS. X-Pod for Horizon View

26 Below are graphs of the Total ISE System IO (Write and Read) in the 3x test runs (500, 1000, and 1500). Write IO can be clearly seen dominating the workload mix. Total IOPS values on the 1,500 user test were observed to regularly occur above 10,000 IOPS at the height of the login phase of the run. Figure 17 - Login Phase and Steady State For 500, 1000, And 1500 Users It was observed that write IOPS were higher in the login phase of the test run. The steady state Login VSI tests required the least amount of performance (IOPS) from the ISE 740 hybrid storage array. Figure 18 - Write Latency for 500, 1000, and 1500 User Tests X-Pod for Horizon View

27 Figure 17 shows that the ISE 740 array performed all of the write operations under 10ms, with 99% of all operations well under 5ms during the 1500 user tests. There are 3x periodic increases in write latency at the end of the 1000 user test due to Storage vmotion operations, which were included in the 1000 series graph to show the effect of this operation during testing. Values of below 1ms are reported as 0ms, and as such the 500x desktop series can be thought of as having no observable latency over 1ms for the testing period. Read latency is also an important measure of system performance, and increases in read latency were observed as the system load increased (as would be expected). The highest values were observed in the 1500 desktop series, with 95% of the average values below 10ms. The Storage vmotion operations (1000) had the most impact to the read latency values; however, this did not appear to be enough to impact the VSImax score significantly. The vast majority of the read latency value for the 500 desktop test were below 1ms (0 values). Figure 19 - Read Latency for 500, 1000, and 1500 User Tests Read latency is the value that will react first when increasing load on the storage system. However, the amount of read IOPS comprises a small percentage of the overall workload. The ISE 740 hybrid storage array demonstrated that it was able to satisfy the Login VSI workloads up to the point of saturation (100% utilization) for the Cisco UCS CPU resources of the 15 Blade servers. The Login phase of the test scenario generated up to twice as many IOPS than the main Login VSI medium workload. If large volumes of users are logging in/out of the environment concurrently, storage performance will play an important role. Virtual Desktop Deploy Operations Deploy operations are something that every environment must go through. Whether performing the initial creation of the desktops or performing a recompose operation, this process replicates copies of the gold image to the various datastores that will contain the desktops and prepares the operating system for use. X-Pod for Horizon View

28 This operation had the least impact to Cisco UCS server resources and was mainly limited by the rate at which desktops were deployed by Horizon View. CPU utilization rates still reached 40%, and memory utilization stabilized at 80% by the end of the process. Consideration should be given to the CPU and memory increase in the event that this process occurs during production hours. The deploy operation performance requirements are biased towards storage write IOPS, at just over 60% of the workload. Average values for total storage IOPS demand were seen to regularly approach 20,000 IOPS, with high values approaching 25,000 IOPS. This process operates over all of the VM active data set size and generates I/O across all of the new desktop capacity. Figure 20 - ISE 740, Total Read and Write IOPS During Deploy Operation Response time is also an important measure to examine when performing system stress testing. The deploy process is one example of what a virtualization administrator may conduct to proof out storage systems being proposed for VDI deployments. The ISE 740 shows excellent read and write response for this workload, with the majority of the write and read latency values below 2ms and 4ms, respectively. The reduction in read and write latency seen at the beginning of the test run is due to the ISE management of data to SSD in real time. This is Continuous Adaptive Data Placement (CADP) in action, as it learns the workload and optimizes for best performance automatically. Figure 21 - Read and Write Latency During Deploy Operation Smaller pool sizes would result in less load being placed on the Cisco UCS CPU and memory resources and would have a significant reduction in the amount of storage IOPS. In relation to the main Login VSI workload, the deploy operation required greater than 200% more storage performance than the medium X-Pod for Horizon View

29 workload. Reduced desktop pools sizes should be considered for smaller environments that do not require deploying or recomposing large numbers of desktops (>500). Over the course of the testing, the desktop creation process was run several times and provided invaluable information on the impact of this operation. The ISE storage array was able to satisfy this workload with excellent response times. Larger numbers of desktops per pool will result in large numbers of deploy operations happening with maintenance operations. This may be completely acceptable if this entire configuration is viewed as a single POD, and there are multiple PODs in the solution. Smaller environments, where this configuration would represent the entire solution, should consider using smaller desktop pool sizes to better control the impact of these operations on production users. Virtual Desktop Boot Storm The virtual machine boot process was the most taxing on the CPU utilization. The figure below shows the processor and memory utilization of a single blade server during this process. CPU utilization reaches saturation (100%) as the different pools of desktops are booted. Limiting the desktop pool sizes should be considered, as this can limit the impact and duration of the event if an entire desktop pool needs to be booted or rebooted. The boot process varied from 100% write to 100% read over the duration of the test. Initially, there is a large write workload, which then changes to mostly read, and then another write component resurfaces toward the end of each group of desktops booted. Write IOPS during this period were observed to reach above 30,000 IOPS and read activity past 40,000 IOPS. This increase in performance is due to the adapting caching of the ISE controller hardware and the automatic real-time management of data between SSD and HDD (CADP). Figure 22 - Total Read and Write IOPS During Boot Storm In total, the ISE 740 hybrid storage array regularly reached levels over 40,000 IOPS during the boot process and transferred data at over 600 MB/sec. X-Pod for Horizon View

30 Figure 23 - Total IOPS and MB/s for Boot Storm Response times of the ISE 740 were well within what would be considered normal for database operations, proving that the ISE was not approaching any limit in performance for this operation. Figure 24 - Read and Write Latency During Boot Storm Boot storms are traditionally extremely difficult for storage systems to keep up with. The broad range of read vs. write requirements, while requiring high-performance IOPS, are usually where most storage systems have significant issues. In this test, the main limiting factor was the Cisco UCS CPU resources as all servers were pushed to 100% CPU utilization. When planning for numbers of consecutive desktops that can be safely started at the same time, careful attention should be paid to the processor utilization of the ESXi servers after high-performance storage is implemented (such as the ISE 740 hybrid storage array). X-Pod for Horizon View