OSS Presentation Accelerating VDI by Daniel Beveridge
- 1. Accelerate VDI - VSA Based I/O Designs Daniel Beveridge Sr. View Architect Vmware.
- 2. Storage Sizing Conundrum How Much to Buy? The Storage Guessing Game •Save Money •Unsure How Many Users •Unsure How Many Users •Overestimate ramp rate •Misunderstand IOPS/user •Misjudge process hurdles •Unsure of Project timeline •Use of Legacy storage •Spend on Space vs. IOPS
- 3. The Result – Declining User Experience
- 4. Workload basics: Comparing Server to Desktop Server Desktop Diversity on cluster Homogeneous on cluster Peaks randomized across apps Synchronized peaks App driven I/O User driven I/O Maintenance is manual or targeted Maintenance is often synchronized on many machines Batch-oriented Real-time oriented Performance fluctuations have low to Performance fluctuations have high moderate visibility and impact visibility and real-time impact Low to moderate workload volatility (peak High workload volatility (peak > 50x avg) < 10x avg) Key differences between workloads • Vastly different volatility • Users react in real-time to performance fluctuations • Severity of reaction to performance fluctuations
- 5. Volatility Examples – High and Low High Volatility (VDI) Low Volatility (Server)
- 6. VDI Volatility – Logon IOPS * Graph by Jim Moyle
- 7. Windows 7 Implications • Up to 40% more IOPS than XP during select operations such as boot. • Boot operations consume about 100 IOPS on average • Select optimizations are possible to brings IOPS consumption more inline with XP • Most customers will use a Win 7 build that is not optimized. Sizing estimates should not assume Win 7 image is tuned properly. • Win 7 deployments need the additional IOPS provided by VSA accelerated storage designs.
- 9. Virtual Storage Appliance (VSA) based Design w/View 4.5
- 10. Why VSA Storage for VDI? • Very Flexible ‘hardware’ configuration • Drivers taken care of by VMware HCL • Low latency I/O delivery to desktops • Limit need for costly networking equipment • Handles random writes very well – VDI’s most challenging load pattern • It’s software – easy to upgrade, re-tune using automation tools.
- 11. Benefits of VSA based Designs Benefits Importance Inexpensive Storage Cost per User drops High IOPS / User (30k+) Great User Experience Burst Isolation Equitable Burst Access Low Latency (1/10ms) Great User Experience Minimizes Network Traffic Lowers risk of bottlenecks Uniform IOPS/User at any Eliminates declining user Scale experience on ramp-up. Scalable Write-Cache Great User Experience – fast writes at any scale Reduce random writes to SSD ACK from RAM for stateless
- 12. VDI – Two Models • Stateless – Users are decoupled from VMs – Most performance benefits – Most operational savings – Most amenable to a purely local disk design – Least expensive, but may require re-working of IT processes leading to implementation challenges • Persistent – Users have their own VM – Uses much more storage – Desktop must be ‘protected’ against data-corruption – Writes must be treated conservatively – limiting performance – More challenging for VSA designs
- 13. Stateless VDI - Benefits General Benefits Benefits of Stateless w/ZFS • Zero risk of desktop corruption • Ability to maximize fast random • Ability to Maximize limited local disk writes (ACK from Ram) – best user resources experience for random writes • Efficient two drive bay config (1 SSD, 1 • Lends itself to a local disk design HDD) • Keep all I/O inside ESX host. • Ability to maximize fast random writes • Enables use of View Composer (ACK from Ram) • Avoid Windows Configuration drift • In-line compression expand disk • Limit support calls – standard image bandwidth 40% for everyone • L2ARC leverages modest SSD without • Faster support resolution times running out of space • Reduce # of VDI machines powered • Excellent caching with ARC on in datacenter • Lowest cost per user
- 14. VDI Design Options – Stateless VDI Design Hints • Create VMDK for L2ARC on SSD ZFS VSA • Disable ZIL – not needed • Publish volume via NFS to ESX • 8-12GB memory for VSA • CPU / memory reservation for VSA • Set pool policy to ‘refresh’ at logoff Local Disks Sample 64GB SSD Format SSD • 15GB for View’s Composer parent image • 40GB VMDK for L2ARC HDD
- 15. Persistent VDI - Benefits General Benefits Benefits of Stateless w/ZFS • Ability to Maximize limited • Least Operational upheaval when local disk resources migrating from PC model • Efficient two drive bay config • Supports user-installed (1 SSD, 1 HDD) applications • Ability to maximize fast • Works better for ‘offline-VDI’ random writes (ACK from ZIL) feature of VMware View • In-line compression expand disk bandwidth 40% • In-line De-Dupe – 2.5x data reduction • L2ARC leverages modest SSD without running out of space • Excellent caching with ARC
- 16. VDI Design Options – Persistent VDI All Local Design Design Hints • Create VSA’s VMDK ‘data disk’ on SAN/NAS via iSCSI of FC • Place ‘data disk’ ZVOL on local HDD in Raidz mirror • Create VMDK for L2ARC on SSD • Place ZIL (SLOG) on SSD • Publish volume via NFS to ESX • 12-16GB memory for VSA • Assign CPU / memory reservation ZFS VSA • Turn on De-Dupe and Compression • Sample 64GB SSD Config: 8GB ZIL, 50GB L2ARC Benefits / Drawbacks • VDI clients get fast local commits on write due to SSD based ZIL (ZFS intent Log) • VSA can 50 write IOPS per user for 150+ users with 1 SLC based SSD as ZIL, and 100+ read IOPS per user. • Inline De-Dupe & Compression reduce I/O from VSA HDD(s) by 80% - better use of limited spindles. ZIL Approximately 30K+ IOPS possible from VSA SSD • • Longer RTO in failure – Host rebuild may be needed L2ARC ZVOL HDD
- 17. VDI Design Options – Persistent VDI Hybrid Design Design Hints SAN/NAS • Create VSA’s VMDK ‘data disk’ on SAN/NAS via iSCSI or FC • Place VSA boot drive on SAN/NAS • Create VMDK for L2ARC on SSD • Place ZIL (SLOG) on SSD iSCSI or FC attached VMDK • Publish volume via NFS to ESX • 12-16GB memory for VSA ZFS VSA • Assign CPU / memory reservation • Sample 64GB SSD Config: 8GB ZIL, 50GB L2ARC Benefits • High # of low-latency random write IOPS possible from VSA with local ZIL, 30K+ Read IOPS. • VDI clients get fast local commits on which makes a great user experience • Inline De-Dupe & Compression reduce I/O from VSA to SAN/NAS by 80% ZIL • Data safer in Host failure scenario – Shorter RTO SSD • May need to move SSD to new ESX host in failure L2ARC scenario
- 18. Isn’t this all a bit Complex? • Different and intricate ZFS configs base on VDI pool type – ZIL status, size, L2ARC size, Raid-z status, De-Dupe/Compress??, hybrid VSA/Central config? • Sizing & Tuning Complexity – How much memory/CPU and other vSphere tuning parameters are needed for each option? – CPU/IO/Mem reservations?, type of VMDK (thin, fat, eager zero?), – iSCSI or NFS mount to ESX? • How does your OS image impact these tunings? – Are there wasteful services running ? – Did a new application change your I/O load levels? • Wow – there’s a lot of steps involved to set this all up? – I feel Carpel tunnel coming on!! I want to use this cool technology – how can it be easier?
- 19. Help is on the way The Nexenta VDI Solution • Setup a whole vSphere cluster – VSA & VDI together in one place • Using ‘provisioning templates’, just answer some simple questions and the rest is taken care of for you. • Pairs ideal ZFS tunings for each VMware View Pool type • Calculates correct resource levels for VSA based on user density targets. • Optimizes VDI density based on IOPS targets you can set. • Runs built-in Micro-Benchmarks to validate performance • Deploy View with confidence and ease – removes the tedious steps and assures an optimized solution.
- 20. Design Principles for Maximizing VDI Benefits 1. Divide Use Cases into Stateless vs. Persistent • Leverage VMware tools like ThinApp to maximize users who qualify for stateless designs – use local VSA storage for these users. • Use VMware’s Virtual Profile tool to keep data out of VM 2. Push as much I/O delivery upstream to VSA as possible • Maximize value of centralized storage for high value IO (offload junk IOPS). • Hybrid VSA/Central storage design for persistent VDI 3. Integration with VMware View’s Tiered Storage • Leverage centralized SSD or local replica for shared OS image when using View Composer • Maximize caching of VSA(s) for unique user files. • Place View’s ‘disposable disk’ on ZVOL with ZIL disabled for persistent VDI VMs • Reduce Logon Storm I/O & data risk with our Virtual Profile feature
- 21. Summary • ZFS features are very well suited to implementing VSA designs for VDI • Key Features include: ZIL, L2ARC, De-Dupe, Compression, ARC caching, Raid-z. • Hybrid VDI design combining SAN/NAS with VSA accelerates I/O, lowers cost • A scale-out VSA storage model to augment SAN/NAS can lower risk of under/over spending by transforming CAPEX into incremental spend. • Use centralized storage for important data, high value I/O and persistent VDI designs. Use local storage for disposable data & non-persistent VDI Last but NOT Least: Look for Nexenta’s exciting new VDI provisioning solution to dramatically simplify the deployment of VSA storage and VMware View desktops (Q1 2012?)