2. Defining Software Defined Storage
• We must resist software defined
washing
• My Criteria:
– Must run on x86 server hardware
• No custom ASICs
• NVRAM allowed
– Can be sold as software or
Appliance
• “Wrapped in tin”
3. Software Defined Storage is not New
• Novell NetWare turned PC AT into file server
– Software defined NAS
• I wrote “How to BuildYour Own iSCSI Array” in
2006
• Faster x86 processors are driving SDS
– In performance
– In scale
– In functionality
4. Software Defined Storage Taxonomy
• Storage SoftwareDefinus OldSchoolus
– Runs on standard server
– Publishes block or file
• Storage SoftwareDefinusVirtualum
– The classicVSA
– Like OldSchoolus but runs in aVM
• Storage SoftwareDefinusVirtucalis Scaleoutus
– Also known as ServerSAN
– Pools storage across hypervisor hosts
5. Old School SDS
• Software runs underWindows or Linux
• Publish storage as iSCSI or file
• Standard RAID
– Could require HW controller
• Synchronous replication with failover
6. Selected Classic SDS
• FreeNAS/OpenNAS Etc.
– Open source ZFS based NAS/iSCSI
• Include SSD caching
• Open-E DSS
– Open source assemblage w/support
• NexentaStor
– Commercial ZFS
• Supports shared storage
• StarWind
– Windows based iSCSI target w/SSD caching
7. Wrap in Tin?
• Most next generation storage arrays are SDS
Nimble Storage Fusion-IO IOcontrol (Nexgen)
Tintri Tegile
• Why?
– Qualification and support
– Margins
– Channel issues
– Customer preferences
8. OrVirtualize The Servers
• Creating aVirtual Storage Appliance (VSA)
• VSAs great solution for ROBO, SMB
• Local storage, may require RAID controller
• Publish as iSCSI or NFS
• Example: Stormagic
– Basic iSCSIVSA 2 nodes $2500
9. Why Converge Storage and Compute?
• Makes corporate data center like cloud
– A good or bad thing
• Storage array slot and SAN costs
– Generally higher than the disk drive that plugs in
• Server slots are already paid for
• Political, management issues
– Moves storage to server team
10. Enter The ServerSAN
• Scale-out whereVSAs are
fail-over clusters
– Storage across n hypervisor
hosts to form one pool
– Maximum cluster sizes 8-32
nodes
• Use SSD as cache or tier
• Can be software or
hyperconvirged servers
11. ServerSAN architecture differentiators
• Data protection model
– Per node RAID?
– N-way replication
– Network RAID?
• Flash usage:
– Write through or write back cache
– SubLUN tiering
• Prioritization/Storage QoS
• Data locality
• Data reduction
• Snapshots and cloning
12. Hyperconvirged Systems
• Nutanix
– Derived from Google File System
– 4 nodes/block
– Multi-hypervisor
– Storage for cluster only
• Simplivity
– Dedupe and backup to the cloud
– Storage available to other servers
– 2u Servers
• Both have compute and storage heavy models
• Pivot3 for VDI only
• ScaleComputing KVM based for SMBs
13. Vmware’sVSAN
• SSD as read/write cache
• N-way replication (no local RAID)
– Default 2 copies, requires 3 nodes
– 3 copies requires 5 nodes (my recommendation)
• Scales to 32 nodes
• Runs directly in hypervisor kernel
• Storage only available to cluster members
• Relies on vSphere snapshots, replication
• License $2495/CPU or $50/VDI image
14. Software Only ServerSANs
• HP StoreVirtual (Lefthand)
– Sub-LUN tiering for SSDs
– iSCSI system scales to 10 nodes
– Data Protection
• Per Node RAID
• 2-4 way replication or network RAID 5 or 6
• Maxta Storage Platform
– Data deduplication and compression
– Metadata based snapshots
– Data integrity via hashes/scrubbing
– Data locality
15. More Software ServerSANs
• EMC ScaleIO
– Extreme scale-out to 100s of nodes
– Multi-hypervisor
• Kernel modules for KVM, XEN, Hyper-V
– Multiple storage pools
– Some QoS
– Metadata snaps and clones
• Sanbolic Melio
– Evolved from clustered file system
– Perhaps the most mature
16. ServerSANs and Server Form Factors
• Mainstream
– 1u servers offer limited storage
– 2u servers the sweet spot
• 6-24 drive bays for both SSDs and HDDs
• Slots for 10Gbps Ethernet and PCIe SSDs
• Blades unsuitable
– 1-2 SFF disk bays
– Mezzanine PCIe SSDs generally >$8000
• High density servers can work
17. Challenges to SDS
• Purchasing politics and budgets
– Everyone likes to point at where their money
went, especially storage guys
– So who’s budget
• Easier if savings big enough that storage+compute now
costs ≤ storage or compute
– VDI can be camel’s nose because it needs
dedicated infrastructure anyway
18. Operational Challenges
• Storage guys are paranoid for good reason
– Storage is persistent and so are storage mistakes
• Server guys are less paranoid
• VSAN with default 2 copies
– VMware admin takes 1 server offline
– The 1 disk with data fails.
19. ServerSAN vs Dedicated Storage
• VMware’s benchmark config
– 32 nodes
– 400GB SSD and 7 1TB drives each
– VSAN cost ~$11,000/server for 2.3TB useable (73TB
total)
• Best of breed dedicated storage
– Tintri T650
• 33.6TB usable $160,000
• ~100,000 real IOPS
• PerVM snaps and replication