3. www.rackspace.com
Requirements
3
• Durability
– How much failure can I experience before data loss?
• Availability
– How much failure can I experience before data unavailable?
– For how long?
• Scalability
– Capacity Today? Tomorrow? Next Year?
– Hot/Cold Ratios
•Performance
– Realistic expectations apply!
4. www.rackspace.com
What exactly are you storing?
4
• Purpose and Value
– Mission Critical, cannot run the business?
– Second Class Services, just inconvenient?
– Dog Pictures?
– Computationally expensive to reproduce?
– Already Compressed or Encrypted?
•Access Patterns
– Random, Highly Transactional?
– Sequential, Streaming?
– Mostly Reads?
– Mostly Writes?
•Does your data age?
– How much of your dataset is “hot” or “cold”?
7. www.rackspace.com
Scale-Out Object Storage
7
• Objects replicated between nodes
• Proxy Layer for load-balancing
• Proven at scale, easy to grow
• Can’t mount file systems or block volumes
• What if my application doesn’t “speak” Object?
• Eventual Consistency
8. www.rackspace.com
LVM / iSCSI
8
• Server(s) with local disks carved up by LVM
– Disk redundancy built in using RAID or Volume Mirroring
• Block Vols presented over Ethernet (iSCSI)
• No node failure protection
– Server fails, volumes fail
• Least Redundancy
• Potential for Hot Spots
• Lowest Cost*
• Easy to Scale
9. www.rackspace.com
Dual Controller Storage Arrays
9
• Typically Termed “Mid-Range” ~50-1000 Disks
• Block Vols presented via iSCSI and FC Protocols
• Controller and Disk Redundancy
• Support for aggregated Disk Pools
• Support for auto-tiering of data
• Snapshot/Clone Capabilities
• Typically 5 9’s Availability
• Limited Scalability
• Good Performance*
10. www.rackspace.com
Multi-Controller Storage Arrays
10
• Typically Termed “High End” ~200 to 3000 Disks
• Block Vols presented via iSCSI and FC Protocols
• Meshed Controller Subsystems
• Support for large aggregated disk pools
• Support for auto-tiering of data
• Snapshot/Clone Capabilities
• Redundant Everything
– 6 9’s and greater availability
• Better Scalability
• Higher Performance*
11. www.rackspace.com
All Flash Arrays
11
• Block Vols presented via iSCSI and FC Protocols
• Controller and Disk Redundancy
• Inline Dedupe and Compression
• Meshed Controller Subsystems
• Snapshot/Clone Capabilities
• Typically 5 9’s Availability
• Lower Scalability
• Highest Performance*
12. www.rackspace.com
Scale-Out / Server SAN
12
• Block Vols presented via iSCSI and Proprietary Protocols
• Typically Commodity Hardware
– x86, 10GbE (sometimes InfiniBand)
• Many nodes, 10’s to 100’s
• Redundancy achieved by spreading data across nodes
– Sub-Object Replication
– Erasure Coding
• Greatest Scaling
• Good Performance*
14. www.rackspace.com
Storage Networking
14
• How do I reach my persistently stored data?
– Ethernet?
– Fibre Channel?
– InfiniBand?
• Any and all Fabrics must be redundant
• Use FC exclusively for Storage, but what about other services?
• Use Ethernet and share with other services?