DVX From PAIN to INSANE- Changing the VDI game with Server Powered Storage Patrick Sliney Principal Systems Engineer @PJSliney
But first, a quick audience survey Who is hosting their applications in Trapp’s App Cloud? I love it! I can finally focus on my core business! I don’t trust my data in the cloud Trapp has a cloud? My VDI project: Is a raging success Is in progress Is in planning / research phase What’s VDI? Compared to last year, I am supporting More users More devices More applications My IT needs are shrinking
But first, a quick audience survey Who’s running these Hypervisors in production? Vmware ESXi Microsoft Hyper-V KVM / Openstack Hypervisors are for hippies- Server huggers FTW! What’s your x86 server platform? Mostly blade servers Mostly chassis servers That x86 stuff will never catch on! MIPS 4ever! How are your servers attached to storage? 10 Gb Ethernet 1 Gb Ethernet Fibre Channel DAS Fibre Channel over Token Ring
The Operational Thesis of Open Converged Array based storage architectures are not addressing the dynamic shift in application workloads being thrust into the hypervisor. Scaling these workloads has exposed critical dependencies that traditional or hyperconverged storage architectures are struggling to support. By driving the application data into the hypervisor hosts, you merge it with the application run-time state and convert to completely localized I/O operations. Server Powered Storage advantages are profound.
Ecosystem Partners VMware Unidesk Citrix Login VSI Horizon View Proven Storage for VDI (in progress) Unidesk http://www.datrium.com/resources/nbmc-success-story/ http://www.datrium.com/resources/abacus-group-success-story/ Citrix XenDesktop Login VSI Tested with Login VSI We are actively working with VMware on both Reference Architecture (RA) and Proven Storage (1000 user) validations and marketing efforts. Unidesk has already shown promise within our existing customer base Citrix works as well with better handling of MCS style deployments on DVX Validation is done internally (and soon to be externally) with Login VSI testing tool framework
Datrium Architecture – Elastic Primary Storage Example ~100 desktops per host 20 core / 256GB Mulitple hosts per VDI cluster 4-12 Over 30TB of usable storage Durable data (with no reduction) Local reads from host flash Reduced latency Data reduction – local / global Host and NetShelf – 6-10x? 10GB
Datrium Benefits for VDI Simpler storage management Single Datastore presented to the infrastructure Data reduction leverages resources throughout environment Better performance isolation Hosts define the performance level Dependency on storage is powered by the host Faster local reads from flash Boot storms (per host) with less contention Some provisioning methods can leverage VAAI clones More flexible deployment alternatives Server selection is not “locked in” Mix and match resources to meet desktop pool requirements
Case Study: Large Oregon Hospital System Scaling: Existing Workload ESX Servers 16 Read Cache 337TB Virtual Desktops 25,040 Speed 0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 |____|____|____|____|____|____|____|____|____|____|____|____|____|____|____|____|____| 480,000 IOP/s 10GbE ESX Servers 3 De-Dup/ Compression Read Cache Cache Hit Ratio Virtual Desktops Speed 0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 |____|____|____|____|____|____|____|____|____|____|____|____|____|____|____|____|____| Single NetShelf 7.4X 44.4TB 100% 658 12,435 IOP/s Server Config Make = HP G8 CPU = 20 Cores SSD = 2X Samsung 850 Pro 1TB ESX = 5.5u2
Open Converged for VMware 10GbE Existing Servers: - Add SSDs 10GbE Integrated: - Bundles 10GbE Custom: Bundled / Existing EASY
Datrium DNA investors Brian Biles, CEO @DDUP: Founder, VP Products Hugo Patterson, CTO @DDUP: 1st CTO, EMC Fellow Sazzala Reddy @DDUP: 2nd CTO, EMC Distinguished Engineer Boris Weissman @VMW: Principal Engineer, 11 years Ganesh Venkitachalam @VMW: Principal Engineer, 12 years NEA and Lightspeed Venture Partners INDIVIDUALS: FORMERLY: Frank Slootman Dr Kai Li Diane Greene Mendel Rosenblum Edouard Bugnion CEO, Data Domain Founder, Data Domain Founder, VMware investors
So… What’s a Datrium? Faster than All Flash Arrays Half the $/GB of Hybrid Arrays More Flexible than Hyperconverged Scale performance and capacity independently
How about a quick whiteboard??
Datrium DVX Components You Have: ESX Hosts Any x86 Server Blades / Rack DVX Hyperdriver VIB installs on ESX Hosts Storage Services: Hyperdriver on ESX Datastore: NFS presented to all hosts READS: All from local flash in each host WRITES: to Netshelf, NVRAM, full HA Dedupe: Both in cache & global on Netshelf You Add: Flash Commodity Abundant Speed from Hosts: Add Hosts, goes faster 10GB DVX Netshelf: We ship, partner installs We Bring: Durable Capacity Appliance 2RU, 100TB+
Datrium DVX Components RAID Controllers Hosts Flash Any host with at least 2 drive slots and 4 CPU cores. One core for Hyperdriver, additional cores for performance. Ideal for Cisco UCS and blade servers. Any from Server Vendor Always configure single-disk RAID-0, or JBOD if no RAID controller First choice Flash recommended by the server vendor Safe Choices – anything on VMWare HCL Latest list can be found in the release notes
Open Converged for VMware 10GbE Existing Servers: - Add SSDs 10GbE Integrated: - Bundles 10GbE Custom: Bundled / Existing EASY
Manage Storage Like VMs Performance Upgrades: Host Only VM HA Event? Boot On Alternate Host Rapid Re-Cache Upload No upload for sequences already in use Layout for flash availability: DiESL Distributed Execution, Shared Logs Sequential, unblended Heat-informed prefetching Consistent Space Reclamation / Sweeping Standard vMotion for SLA adjustments If local host resource-limited, after vMotion: Sequential cache upload from NetShelf Read from origin cache until destination is warm 1 . 2
Unlocking ESX Performance with SPS Typically there are empty disk slots available in rackmount or blade servers. example: cache needed: average number of VMs / host = 40 * average size of VM = 100GB = 4TB cache needed to provide 100% cached reads 8TB RAW Flash with capacity efficiency factor 3X (compression + deduplication) = 24TB cache / host available 6X cache over-subscription possible – maintained as your cluster grows out 4TB hypervisor host 4TB hypervisor host 4TB hypervisor host 4TB hypervisor host 4TB hypervisor host 4TB hypervisor host 4TB hypervisor host 4TB hypervisor host 4TB hypervisor host 4TB hypervisor host
Case Study: Large NY Email Hosting Firm Scaling: Existing Workload ESX Servers 12 Read Cache 50TB VMs 900 Speed 0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 |____|____|____|____|____|____|____|____|____|____|____|____|____|____|____|____|____| 360,000 IOP/s 10GbE ESX Servers 4 De-Dup/ Compression 5.2X Read Cache Cache Hit Ratio VMs Speed 0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 |____|____|____|____|____|____|____|____|____|____|____|____|____|____|____|____|____| Single NetShelf 16.6TB 100% 171 3824 IOP/s Server Config Make = HP G8 CPU = 20 Cores SSD = 2X Samsung 850 Pro 1TB ESX = 5.5u2
Datrium Storage No storage training needed, VM Centric Disk @ .50c /GB - Flash @ .30c /GB More cores, more hosts = more speed X1000 IOP/s 0 100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 |____|____|____|____|____|____|____|____|____|____|____|____|____|____|____|____|____|
Manage Storage Like VMs Completely Integrated vSphere Presence No other UI to Learn or Access Manage at DVX, Host, and VM levels Summary, Real-time & Historical Views Per Host, VM, and DVX At a Glance Situational Knowledge Real-time Key Performance Metrics Event and Task Oriented Monitor and Manage Capacity and Flash Resources Headroom Available, Data Reduction Available and Projected Stats Simplify
Why Datrium Simple Concepts: No more islands of SAN Manage Hosts and Applications, not arrays Leverage & augment existing infrastructure Lower TCO Evolve to faster more efficient storage vs rip replace Faster Cheaper Easier
Fundamentally different