1 WWW.FUSIONIO.COM WHITE PAPER
WHITE PAPER Executive Summary Fusion iovdi is the first desktop- aware solution to virtual desktop infrastructure. Its software- defined approach uniquely combines the economics and management of server- side flash with the data protection and virtualization benefits of shared storage to solve the performance, cost, and management hurdles of virtual desktop infrastructure. Fusion- io conducted tests comparing an identical server attached to shared storage with and without the iovdi acceleration software and iomemory server flash. The results were dramatic: SAN/NAS Write workload offloaded up to 80% with Write Vectoring. iovdi reduces shared storage performance dependencies by differentiating between persistent and non- persistent desktop writes. The iovdi filter driver identifies non- persistent desktop write activity, such as pagefile.sys writes, and directs it to fast server- side flash that resides close to the virtual desktop application, making desktops more responsive. Persistent data write activity is vectored to shared storage to support vmotion, HA, DRS, and SRM. 6X Improvement in boot times with Transparent File Sharing. iovdi reduces shared storage performance dependencies even further by allowing desktop images to simultaneously share common requested files from fast server flash, speeding both individual desktop boot times and large- scale desktop startups such as during a VMware HA restart. Reduced VRAM Provisioning by 88%. iovdi alleviates memory contention by presenting flash as a natural extension of RAM for page swaps for example. This allows administrators to reduce VRAM per VM without impacting performance and frees CPU cycles for use by VMs. Consistent 50ms desktop response times. iovdi delivers consistent 50 ms desktop response times by eliminating most reads and writes to shared storage, thus reducing dependencies on storage or storage area network fabric limitations. The results show that the iovdi solution dramatically reduces the system s need for shared storage resources, while more efficiently using host resources to enable fully persistent desktops at low cost, with a user experience as good or better than traditional laptops with SSDs. 2 WWW.FUSIONIO.COM
Test System Overview Figure 1. Simplified Test Bed. SERVER HARDWARE SuperMicro Twin with 2 Servers for 800 total virtual desktops AMD 6376 256G 1.2 TB PCIe iomemory Flash per ESXi host 1 GbE SERVER SOFTWARE ESXi 5.0 build 702118 iovdi server- side driver Fusion iomemory VSL driver VMware Horizon View 5.2 64bit Windows 7 desktops with VDI enhancements iovdi guest driver STORAGE HARDWARE NTAP FAS 6030 Single controller 1 GbE iscsi SAN interface 2.0T Capacity with 48 5400 rpm NL SAS drives 3 WWW.FUSIONIO.COM
Benchmark Methodology SETUP The VDI environment was created using a standard installation of vsphere 5.0 on the host servers. The VSL and iovdi host components were then installed. NFS was used to host the Windows 7 virtual desktop disk images. The ESXi hosts were connected to shared storage via a 1Gbit Ethernet network switch. The ESXi hosts also had the NetApp RCU plugin installed. The virtual desktops and benchmark infrastructure all resided on a dedicated 10GbE private network with Active Directory, DNS and DHCP services available to all the virtual machines. An Active Directory Domain Controller (DC) was deployed on a Windows Server 2008 R2 guest virtual machine. The guest had the latest VMware tools installed. The VMware tools were configured to synchronize the guest clock with the host. The Login VSI Active Directory setup scripts were run on the DC to create the computer and user accounts required for the benchmark. All virtual desktops were cloned from the same Windows 7 Ultimate 64- bit base image. The base image had only one vnic on the private 1GbE network. The virtual disk size was 25GB and thinly provisioned. The latest VMware Tools were installed, along with the latest Windows updates. The iovdi guest component was installed to enable Transparent File Sharing (TFS). Installation consisted of running an MSI package in the guest base image before provisioning the pool. For these tests, this was a one- time installation. However, installation can also be automated post- provisioning to an existing desktop pool. For comprehensive information on optimizing Windows 7 for VDI use with VMware View, see the following document: http://www.vmware.com/files/pdf/vmware- View- OptimizationGuideWindows7- EN.pdf The batch script in Appendix A of the VMware View Optimization Guide was run on the Windows 7 base image. In addition, the balloon driver was disabled and VMX swapping was disabled. A vcenter Server Appliance 5.0 (vcsa) was installed to manage the virtual desktops. After the base image was prepared, 800 virtual desktops were provisioned, 200 on each ESXi host. The NetApp Virtual Storage Console (VSC) plugin was installed on the DC and registered with the vcsa. The Rapid Cloning Utility (RCU) functionality of the VSC was used to provision the desktop virtual machines on the NetApp. LOGINVSI LoginVSI was used to measure the workload that desktops created on the storage. The virtual machines for the LoginVSI infrastructure were all based on a Windows Server 2008 R2 guest OS. LoginVSI launchers were connected to the desktops via the 10GbE private network. iovdi can be used with any desktop protocol, including Microsoft Remote Desktop Protocol (RDP), Citrix ICA/HDX protocol, and VMware PCoIP protocol. For this benchmark, the LoginVSI Launchers used RDP as the protocol for initiating login sessions. While each benchmark was running, the number of desktops was increased sequentially from 1 to 800 concurrent desktops. 4 WWW.FUSIONIO.COM
BOOT STORM GENERATION Eight hundred Windows 7 desktops were powered on at the same time without any throttling. Total boot time started from the moment the power on button was clicked until activity on primary storage died down. This happened when all of the VMs were on the login screen. Benchmarks Results This section describes the results of the benchmark. For all the tests, the number of concurrent desktop sessions was ramped from 1 to 800 by sequential logins with a session interval of 10 seconds. In all the charts the x- axis indicates the number of concurrent desktop sessions running the workload. LOGINVSI RESULTS The following charts show the storage workload that desktop logins generated with and without iovdi. Figure 2. iovdi supports double the desktops on less than half the storage resources. When iovdi is enabled there is still plenty of performance left in the storage array, even with 800 desktops running. In contrast, at around 400 desktops, the shared storage can no longer keep up with the I/O workload all disks are fully utilized. Adding performance requires purchasing more expensive spindles and perhaps even a new storage array and sizing for performance is tricky. These results show that iovdi greatly reduces the workload VDI generates on storage, freeing resources for additional desktops or for other enterprise infrastructure storage requirements. 5 WWW.FUSIONIO.COM
READ OFFLOAD iovdi caches reads to local server flash, which accelerates desktop performance while reducing the read workload on storage. To gauge the effectiveness of iovdi s caching on the server, we measured the storage read activity for the test. The chart below shows the results: Figure 3. iovdi nearly eliminates VDI read workload from storage. Without iovdi, the storage read activity is heavy, peaking at around 60 MB/sec before the array is saturated. By contrast, even with the full workload of 800 concurrent desktops, iovdi has cached nearly all read activity, greatly reducing load on storage. 6 WWW.FUSIONIO.COM
WRITE OFFLOAD iovdi offloads writes that do not need to be persisted to storage, such as page file and temporary files used by Windows. To gauge how well iovdi reduced writes on storage, we measured storage write activity for the test. The chart below shows the results: Figure 4. iovdi dramatically reduced write workload on storage. With iovdi, even the small storage array we are using has plenty of headroom at 800 desktops. Without iovdi, the storage array runs out of steam at around 400 desktops. 7 WWW.FUSIONIO.COM
CACHE EFFICIENCY The chart below shows the distribution of reads and writes being served either by iovdi or by the storage array, measured in total MB throughput the workload. As the chart shows, the VDI workload is a perfect fit for the Fusion iovdi features outlined in this document, as nearly all the data traffic is redirected to the server- side flash storage. The benefit to administrators is reclaimed capacity and less stress on the backend stores, providing a more efficient and flexible storage subsystem. The benefit to the virtual desktop user is a high quality desktop image and applications that can be more responsive than those on of a physical laptop with an SSD. Figure 5. iovdi directs local traffic to high- performance flash and persistent writes to storage. 8 WWW.FUSIONIO.COM
USER EXPERIENCE When a virtual desktop experiences significant delays from the I/O subsystem, the user is left waiting. On a physical machine, storage is dedicated to the user and response time can be kept low. In contrast, VDI deployments often fail because the shared storage infrastructure cannot deliver the latencies that users have come to expect from dedicated storage. Figure 6. iovdi delivers a user experience that can exceed physical machines with SSDs. At 300 desktops and with iovdi disabled, the latency of the I/O operations on the host starts climbing dramatically as the shared storage struggles to keep up. As latencies approach 10ms, users experience perceptible delays in their desktop experience and as it goes above 100ms, latency significantly impacts the user experience. With iovdi enabled, latencies are kept consistently low, even with 800 users concurrently accessing the server, I/O latencies and desktop responsiveness can be lower than even a physical PC, resulting in a much better user experience. 9 WWW.FUSIONIO.COM
CPU EFFICIENCY A key constraint for many virtualized environments is CPU. The tests measured the CPU efficiency with and without iovdi, which affects server desktop density. The chart below shows the results: Figure 7. iovdi achieves greater desktop density through efficient CPU utilization. With iovdi enabled, all allocated CPU resources can be fully utilized by all 800 desktops. Note that a more powerful CPU can support even more desktops, but in this test the four 2- socket CPU limited support to 800 desktops. Without IoVDI, the CPU utilization of the hypervisor is gated by the storage infrastructure. At 400 desktops, desktops are waiting for I/O and the VDI solution can only use 50-60% of the CPU allocated to it, effectively wasting nearly half of the server CPU. This inefficiency effectively doubles the licensing costs of the storage- based solution compared to the iovdi solution. 10 WWW.FUSIONIO.COM
BOOT STORM RESILIENCY A common problem in VDI environments is boot storms. When a large number of desktops boot at the same time, (for example, after a power outage or server maintenance) the storage subsystem is often overloaded with I/O requests, resulting in very long boot times and poor user experiences. To measure boot storm resiliency, we examined the I/O activity at shared (primary) storage during a reboot of 800 desktops. The vertical lines mark the point where all desktops have booted and are available for use. Figure 8. iovdi boots virtual desktops 6x faster than the storage- based system. iovdi reduces the impact of boot storms by allowing many virtual desktops to simultaneously share common files. The Transparent File Sharing feature allowed all 800 desktops to fully boot in just 5 minutes. Without iovdi, the reboot of all 800 desktops took over 30 minutes. Fast reboots are also provided for individual desktops. With the pool of all 800 virtual desktops active, any single virtual desktop reboots in just 8-12 seconds compared with over 1 minute without Fusion iovdi software. 11 WWW.FUSIONIO.COM
iovdi Benefit Recap The testing shows that iovdi offers the following benefits: Fast, predictable desktop experience. Thousands of server- side IOPS are available for every desktop user at very low latencies to ensure fast, consistent response times and rapid boots, even at peak hours. Simple scale- out. When your system needs to support more VDI users, you can add iovdi appliances in a building block fashion to seamlessly and dynamically scale with no disruption to the production environment. Easy deployment. iovdi is a server- based solution that is ideally suited to VMware View environments and can be managed without adding personnel or requiring specialized training. Lower capex costs. iovdi reduces the capital expense of VDI deployment costs by efficiently using existing shared storage capacity for persistence, flash resources for performance, and conserving network bandwidth for only writes that need to cross the wire. Full VDI support. iovdi supports advanced VDI functionality, including VMware features such as vstorage API for Array Integration (VAAI). Unparalleled desktop density. In this example, a single 2U iovdi appliance delivers enough performance and storage capacity to power up to 800 virtual desktops. Eliminate infrastructure over- provisioning. Intelligent write vectoring uses storage only for data that must be persisted, while using flash for what it does best provide low- latency performance for fast response times, boot times, and cloning, even under load. Key iovdi Innovations WRITE VECTORING In virtual desktops, many writes (writes to temporary, swap and other transient files) do not need to be persisted to shared storage. Write vectoring transparently and intelligently redirects ~80% of writes to local flash storage, greatly offloading the shared storage infrastructure. TRANSPARENT FILE SHARING (TFS) TFS provides an in- line, file- level deduplication of all desktops hosted on a server. Unlike block- level deduplication solutions, this is a fast- path solution that operates at the application level, which serves I/O directly from the cache. With a typical 80% deduplication ratio for VDI, iovdi makes optimal use of flash capacity, providing sufficient space to support read caching and write vectoring for all desktops on a fully loaded server. TFS also helps eliminate boot and other storms by allowing VMs to access data locally on server- side flash rather than on shared storage. READ CACHE Read cache offloads reads from backend storage to local server- based flash cache. This greatly improves application performance on the desktop by dramatically reducing read latencies. Moreover, it allows backend storage to optimize its performance for writes, providing better overall system behavior. 12 WWW.FUSIONIO.COM
Conclusion Fusion- io conducted tests comparing an identical server attached to shared storage with and without the iovdi acceleration software with iomemory server flash. The results were dramatic: SAN/NAS WRITE offloaded up to 80% with Write Vectoring. iovdi reduces shared storage performance dependencies by differentiating between persistent and non- persistent desktop writes. The iovdi filter driver identifies non- persistent desktop write activity, such as pageswap.sys file updates, which is directed, or vectored, to server- side flash for acceleration close to the virtual desktop application. Persistent data write activity is vectored to shared storage to support vmotion, HA, DRS, and SRM. 6x Improvement in boot times with Transparent File Sharing. iovdi reduces shared storage performance dependencies even further by allowing desktop images to simultaneously share common requested files from fast server flash, speeding both individual desktop boot times and large- scale desktop startups such as during a VMware HA restart. Reduced VRAM Provisioning by 88%. iovdi alleviates memory contention by paging to flash which allows administrators to reduce VRAM per VM without impacting performance and frees CPU cycles for use by VMs. Consistent 50ms desktop response times. iovdi delivered consistent 50ms desktop response times by eliminating most reads and writes to shared storage, thus reducing dependencies on storage or storage area network fabric limitations. The results show that the iovdi solution dramatically reduces the system s need for shared storage resources, while more efficiently using host resources to enable fully persistent desktops at low cost, with a user experience as good or better than traditional laptops, PCs or workstations with SSDs. 2014 Fusion- io, Inc. All rights reserved. iovdi, VSL, and iomemory are trademarks or registered trademarks of Fusion- io, Inc. All other product and company names may be trademarks of the companies with which they are associated. 13 WWW.FUSIONIO.COM