Keynote: Building Tomorrow's Ceph - Ceph Day Frankfurt

Technology

inktank
  • 1. Building Tomorrow's Ceph Sage Weil
  • 2. Research beginnings 9
  • 3. UCSC research grant  “Petascale object storage”  US Dept of Energy: LANL, LLNL, Sandia  Scalability  Reliability  Performance  Raw IO bandwidth, metadata ops/sec  HPC file system workloads  Thousands of clients writing to same file, directory
  • 4. Distributed metadata management  Innovative design  Subtree-based partitioning for locality, efficiency  Dynamically adapt to current workload  Embedded inodes  Prototype simulator in Java (2004)  First line of Ceph code  Summer internship at LLNL  High security national lab environment  Could write anything, as long as it was OSS
  • 5. The rest of Ceph  RADOS – distributed object storage cluster (2005)  EBOFS – local object storage (2004/2006)  CRUSH – hashing for the real world (2005)  Paxos monitors – cluster consensus (2006) → emphasis on consistent, reliable storage → scale by pushing intelligence to the edges → a different but compelling architecture
  • 6.  Click to edit the outline text format  Second Outline Level Third Outline Level  Fourth Outline Level Fifth Outline Level Sixth Outline Level Seventh Outline Level Eighth Outline Level Ninth Outline LevelClick to edit Master text styles
  • 7. Industry black hole  Many large storage vendors  Proprietary solutions that don't scale well  Few open source alternatives (2006)  Very limited scale, or  Limited community and architecture (Lustre)  No enterprise feature sets (snapshots, quotas)  PhD grads all built interesting systems...  ...and then went to work for Netapp, DDN, EMC, Veritas.  They want you, not your project
  • 8. A different path  Change the world with open source  Do what Linux did to Solaris, Irix, Ultrix, etc.  What could go wrong?  License  GPL, BSD...  LGPL: share changes, okay to link to proprietary code  Avoid community un-friendly practices  No dual licensing  No copyright assignment
  • 9. Incubation 17
  • 10. DreamHost!  Move back to Los Angeles, continue hacking  Hired a few developers  Pure development  No deliverables
  • 11. Ambitious feature set  Native Linux kernel client (2007-)  Per-directory snapshots (2008)  Recursive accounting (2008)  Object classes (2009)  librados (2009)  radosgw (2009)  strong authentication (2009)  RBD: rados block device (2010)
  • 12. The kernel client  ceph-fuse was limited, not very fast  Build native Linux kernel implementation  Began attending Linux file system developer events (LSF)  Early words of encouragement from ex-Lustre devs  Engage Linux fs developer community as peer  Eventually merged CephFS client for v2.6.34 (early 2010)  RBD client merged in 2011
  • 13. Part of a larger ecosystem  Ceph need not solve all problems as monolithic stack  Replaced ebofs object file system with btrfs  Same design goals  Robust, well optimized  Kernel-level cache management  Copy-on-write, checksumming, other goodness  Contributed some early functionality  Cloning files  Async snapshots
  • 14. Budding community  #ceph on irc.oftc.net, ceph-devel@vger.kernel.org  Many interested users  A few developers  Many fans  Too unstable for any real deployments  Still mostly focused on right architecture and technical solutions
  • 15. Road to product  DreamHost decides to build an S3-compatible object storage service with Ceph  Stability  Focus on core RADOS, RBD, radosgw  Paying back some technical debt  Build testing automation  Code review!  Expand engineering team
  • 16. The reality  Growing incoming commercial interest  Early attempts from organizations large and small  Difficult to engage with a web hosting company  No means to support commercial deployments  Project needed a company to back it  Fund the engineering effort  Build and test a product  Support users  Bryan built a framework to spin out of DreamHost
  • 17. Launch 26
  • 18. Do it right  How do we build a strong open source company?  How do we build a strong open source community?  Models?  RedHat, Cloudera, MySQL, Canonical, …  Initial funding from DreamHost, Mark Shuttleworth
  • 19. Goals  A stable Ceph release for production deployment  DreamObjects  Lay foundation for widespread adoption  Platform support (Ubuntu, Redhat, SuSE)  Documentation  Build and test infrastructure  Build a sales and support organization  Expand engineering organization
  • 20. Branding  Early decision to engage professional agency  MetaDesign  Terms like  “Brand core”  “Design system”  Keep project and company independent  Inktank != Ceph  The Future of Storage
  • 21.  Click to edit the outline text format Second Outline Level Slick graphics  broken powerpoint template 31
  • 22. Today: adoption 32
  • 23. Traction  Too many production deployments to count  We don't know about most of them!  Too many customers (for me) to count  Expansive partner list  Lots of inbound  Lots of press and buzz
  • 24. Quality  Increased adoption means increased demands on robust testing  Across multiple platforms  Upgrades  Rolling upgrades  Inter-version compatibility
  • 25. Developer community  Significant external contributors  Many full-time contributors outside of Inktank  First-class feature contributions from contributors  Non-Inktank participants in daily stand-ups  External access to build/test lab infrastructure  Common toolset  Github  Email (kernel.org)  IRC (oftc.net)  Linux distros
  • 26. CDS: Ceph Developer Summit  Community process for building project roadmap  100% online  Google hangouts  Wikis  Etherpad  Quarterly  Our 4th CDS next week  Great participation  Ongoing indoctrination of Inktank engineers to open development model
  • 27. Erasure coding  Replication for redundancy is flexible and fast  For larger clusters, it can be expensive  Erasure coded data is hard to modify, but ideal for cold or read-only objects  Will be used directly by radosgw  Coexists with new tiering capability Storage overhead Repair traffic MTTDL (days) 3x replication 3x 1x 2.3 E10 RS (10, 4) 1.4x 10x 3.3 E13 LRC (10, 6, 5) 1.6x 5x 1.2 E15
  • 28. Tiering  Client side caches are great, but only buy so much.  Separate hot and cold data onto different storage devices  Promote hot objects into a faster (e.g., flash-backed) cache pool  Push cold object back into slower (e.g., erasure-coded) base pool  Use bloom filters to track temperature  Common in enterprise solutions; not found in open source scale-out systems → new (with erasure coding) in Firefly release
  • 29. The Future 40
  • 30. Technical roadmap  How do we reach new use-cases and users  How do we better satisfy existing users  How do we ensure Ceph can succeed in enough markets for supporting organizations to thrive  Enough breadth to expand and grow the community  Enough focus to do well
  • 31. Multi-datacenter, geo-replication  Ceph was originally designed for single DC clusters  Synchronous replication  Strong consistency  Growing demand  Enterprise: disaster recovery  ISPs: replication data across sites for locality  Two strategies:  use-case specific: radosgw, RBD  low-level capability in RADOS
  • 32. RGW: Multi-site and async replication  Multi-site, multi-cluster  Regions: east coast, west coast, etc.  Zones: radosgw sub-cluster(s) within a region  Can federate across same or multiple Ceph clusters  Sync user and bucket metadata across regions  Global bucket/user namespace, like S3  Synchronize objects across zones  Within the same region  Across regions  Admin control over which zones are master/slave
  • 33. RBD: block devices  Today: backup capability  Based on block device snapshots  Efficiently mirror changes between consecutive snapshots across clusters  Now supported/orchestrated by OpenStack  Good for coarse synchronization (e.g., hours or days)  Tomorrow: data journaling for async mirroring  Pending blueprint at next week's CDS  Mirror active block device to remote cluster  Possibly with some configurable delay
  • 34. Async replication in RADOS  One implementation to capture multiple use-cases  RBD, CephFS, RGW, … RADOS  A harder problem  Scalable: 1000s OSDs → 1000s of OSDs  Point-in-time consistency  Challenging research problem → Ongoing design discussion among developers
  • 35. CephFS → This is where it all started – let's get there  Today  Stabilization of multi-MDS, directory fragmentation, QA  NFS, CIFS, Hadoop/HDFS bindings complete but not productized  Need  Greater QA investment  Fsck  Snapshots  Amazing community effort (Intel, NUDT and Kylin)  2014 is the year
  • 36. Governance How do we strengthen the project community?  2014 is the year  Recognized project leads  RBD, RGW, RADOS, CephFS, ...  Formalize emerging processes around CDS, community roadmap  External foundation?
  • 37. The larger ecosystem
  • 38. The enterprise How do we pay for all of this?  Support legacy and transitional client/server interfaces  iSCSI, NFS, pNFS, CIFS, S3/Swift  VMWare, Hyper-V  Identify the beachhead use-cases  Earn others later  Single platform – shared storage resource  Bottom-up: earn respect of engineers and admins  Top-down: strong brand and compelling product
  • 39. Why Ceph is the Future of Storage  It is hard to compete with free and open source software  Unbeatable value proposition  Ultimately a more efficient development model  It is hard to manufacture community  Strong foundational architecture  Next-generation protocols, Linux kernel support  Unencumbered by legacy protocols like NFS  Move from client/server to client/cluster  Ongoing paradigm shift  Software defined infrastructure, data center  Widespread demand for open platforms
  • 40.  Click to edit the outline text format Second Outline Level Thank you, and Welcome!
    Please download to view
  • 51
    All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
    Description
    Sage Weil, Founder & CTO, Inktank
    Text
    • 1. Building Tomorrow's Ceph Sage Weil
  • 2. Research beginnings 9
  • 3. UCSC research grant  “Petascale object storage”  US Dept of Energy: LANL, LLNL, Sandia  Scalability  Reliability  Performance  Raw IO bandwidth, metadata ops/sec  HPC file system workloads  Thousands of clients writing to same file, directory
  • 4. Distributed metadata management  Innovative design  Subtree-based partitioning for locality, efficiency  Dynamically adapt to current workload  Embedded inodes  Prototype simulator in Java (2004)  First line of Ceph code  Summer internship at LLNL  High security national lab environment  Could write anything, as long as it was OSS
  • 5. The rest of Ceph  RADOS – distributed object storage cluster (2005)  EBOFS – local object storage (2004/2006)  CRUSH – hashing for the real world (2005)  Paxos monitors – cluster consensus (2006) → emphasis on consistent, reliable storage → scale by pushing intelligence to the edges → a different but compelling architecture
  • 6.  Click to edit the outline text format  Second Outline Level Third Outline Level  Fourth Outline Level Fifth Outline Level Sixth Outline Level Seventh Outline Level Eighth Outline Level Ninth Outline LevelClick to edit Master text styles
  • 7. Industry black hole  Many large storage vendors  Proprietary solutions that don't scale well  Few open source alternatives (2006)  Very limited scale, or  Limited community and architecture (Lustre)  No enterprise feature sets (snapshots, quotas)  PhD grads all built interesting systems...  ...and then went to work for Netapp, DDN, EMC, Veritas.  They want you, not your project
  • 8. A different path  Change the world with open source  Do what Linux did to Solaris, Irix, Ultrix, etc.  What could go wrong?  License  GPL, BSD...  LGPL: share changes, okay to link to proprietary code  Avoid community un-friendly practices  No dual licensing  No copyright assignment
  • 9. Incubation 17
  • 10. DreamHost!  Move back to Los Angeles, continue hacking  Hired a few developers  Pure development  No deliverables
  • 11. Ambitious feature set  Native Linux kernel client (2007-)  Per-directory snapshots (2008)  Recursive accounting (2008)  Object classes (2009)  librados (2009)  radosgw (2009)  strong authentication (2009)  RBD: rados block device (2010)
  • 12. The kernel client  ceph-fuse was limited, not very fast  Build native Linux kernel implementation  Began attending Linux file system developer events (LSF)  Early words of encouragement from ex-Lustre devs  Engage Linux fs developer community as peer  Eventually merged CephFS client for v2.6.34 (early 2010)  RBD client merged in 2011
  • 13. Part of a larger ecosystem  Ceph need not solve all problems as monolithic stack  Replaced ebofs object file system with btrfs  Same design goals  Robust, well optimized  Kernel-level cache management  Copy-on-write, checksumming, other goodness  Contributed some early functionality  Cloning files  Async snapshots
  • 14. Budding community  #ceph on irc.oftc.net, ceph-devel@vger.kernel.org  Many interested users  A few developers  Many fans  Too unstable for any real deployments  Still mostly focused on right architecture and technical solutions
  • 15. Road to product  DreamHost decides to build an S3-compatible object storage service with Ceph  Stability  Focus on core RADOS, RBD, radosgw  Paying back some technical debt  Build testing automation  Code review!  Expand engineering team
  • 16. The reality  Growing incoming commercial interest  Early attempts from organizations large and small  Difficult to engage with a web hosting company  No means to support commercial deployments  Project needed a company to back it  Fund the engineering effort  Build and test a product  Support users  Bryan built a framework to spin out of DreamHost
  • 17. Launch 26
  • 18. Do it right  How do we build a strong open source company?  How do we build a strong open source community?  Models?  RedHat, Cloudera, MySQL, Canonical, …  Initial funding from DreamHost, Mark Shuttleworth
  • 19. Goals  A stable Ceph release for production deployment  DreamObjects  Lay foundation for widespread adoption  Platform support (Ubuntu, Redhat, SuSE)  Documentation  Build and test infrastructure  Build a sales and support organization  Expand engineering organization
  • 20. Branding  Early decision to engage professional agency  MetaDesign  Terms like  “Brand core”  “Design system”  Keep project and company independent  Inktank != Ceph  The Future of Storage
  • 21.  Click to edit the outline text format Second Outline Level Slick graphics  broken powerpoint template 31
  • 22. Today: adoption 32
  • 23. Traction  Too many production deployments to count  We don't know about most of them!  Too many customers (for me) to count  Expansive partner list  Lots of inbound  Lots of press and buzz
  • 24. Quality  Increased adoption means increased demands on robust testing  Across multiple platforms  Upgrades  Rolling upgrades  Inter-version compatibility
  • 25. Developer community  Significant external contributors  Many full-time contributors outside of Inktank  First-class feature contributions from contributors  Non-Inktank participants in daily stand-ups  External access to build/test lab infrastructure  Common toolset  Github  Email (kernel.org)  IRC (oftc.net)  Linux distros
  • 26. CDS: Ceph Developer Summit  Community process for building project roadmap  100% online  Google hangouts  Wikis  Etherpad  Quarterly  Our 4th CDS next week  Great participation  Ongoing indoctrination of Inktank engineers to open development model
  • 27. Erasure coding  Replication for redundancy is flexible and fast  For larger clusters, it can be expensive  Erasure coded data is hard to modify, but ideal for cold or read-only objects  Will be used directly by radosgw  Coexists with new tiering capability Storage overhead Repair traffic MTTDL (days) 3x replication 3x 1x 2.3 E10 RS (10, 4) 1.4x 10x 3.3 E13 LRC (10, 6, 5) 1.6x 5x 1.2 E15
  • 28. Tiering  Client side caches are great, but only buy so much.  Separate hot and cold data onto different storage devices  Promote hot objects into a faster (e.g., flash-backed) cache pool  Push cold object back into slower (e.g., erasure-coded) base pool  Use bloom filters to track temperature  Common in enterprise solutions; not found in open source scale-out systems → new (with erasure coding) in Firefly release
  • 29. The Future 40
  • 30. Technical roadmap  How do we reach new use-cases and users  How do we better satisfy existing users  How do we ensure Ceph can succeed in enough markets for supporting organizations to thrive  Enough breadth to expand and grow the community  Enough focus to do well
  • 31. Multi-datacenter, geo-replication  Ceph was originally designed for single DC clusters  Synchronous replication  Strong consistency  Growing demand  Enterprise: disaster recovery  ISPs: replication data across sites for locality  Two strategies:  use-case specific: radosgw, RBD  low-level capability in RADOS
  • 32. RGW: Multi-site and async replication  Multi-site, multi-cluster  Regions: east coast, west coast, etc.  Zones: radosgw sub-cluster(s) within a region  Can federate across same or multiple Ceph clusters  Sync user and bucket metadata across regions  Global bucket/user namespace, like S3  Synchronize objects across zones  Within the same region  Across regions  Admin control over which zones are master/slave
  • 33. RBD: block devices  Today: backup capability  Based on block device snapshots  Efficiently mirror changes between consecutive snapshots across clusters  Now supported/orchestrated by OpenStack  Good for coarse synchronization (e.g., hours or days)  Tomorrow: data journaling for async mirroring  Pending blueprint at next week's CDS  Mirror active block device to remote cluster  Possibly with some configurable delay
  • 34. Async replication in RADOS  One implementation to capture multiple use-cases  RBD, CephFS, RGW, … RADOS  A harder problem  Scalable: 1000s OSDs → 1000s of OSDs  Point-in-time consistency  Challenging research problem → Ongoing design discussion among developers
  • 35. CephFS → This is where it all started – let's get there  Today  Stabilization of multi-MDS, directory fragmentation, QA  NFS, CIFS, Hadoop/HDFS bindings complete but not productized  Need  Greater QA investment  Fsck  Snapshots  Amazing community effort (Intel, NUDT and Kylin)  2014 is the year
  • 36. Governance How do we strengthen the project community?  2014 is the year  Recognized project leads  RBD, RGW, RADOS, CephFS, ...  Formalize emerging processes around CDS, community roadmap  External foundation?
  • 37. The larger ecosystem
  • 38. The enterprise How do we pay for all of this?  Support legacy and transitional client/server interfaces  iSCSI, NFS, pNFS, CIFS, S3/Swift  VMWare, Hyper-V  Identify the beachhead use-cases  Earn others later  Single platform – shared storage resource  Bottom-up: earn respect of engineers and admins  Top-down: strong brand and compelling product
  • 39. Why Ceph is the Future of Storage  It is hard to compete with free and open source software  Unbeatable value proposition  Ultimately a more efficient development model  It is hard to manufacture community  Strong foundational architecture  Next-generation protocols, Linux kernel support  Unencumbered by legacy protocols like NFS  Move from client/server to client/cluster  Ongoing paradigm shift  Software defined infrastructure, data center  Widespread demand for open platforms
  • 40.  Click to edit the outline text format Second Outline Level Thank you, and Welcome!
  • Comments
    Top