Inception Pack Vol 2: Bizarre premium
•
2 likes•482 views
Cumulus is a filesystem backup utility that uses cloud storage on Amazon S3. It addresses the scalability issues of tape drives for backup. Cumulus segments files to reduce costs from small file sizes and uses sub-file incrementals to only store changed portions of files between snapshots. Evaluation of Cumulus showed it significantly reduced storage costs and backup time compared to traditional methods.
Recommended
More Related Content
What's hot
What's hot (20)
Similar to Inception Pack Vol 2: Bizarre premium
Similar to Inception Pack Vol 2: Bizarre premium (20)
More from The Planning Lab
More from The Planning Lab (14)
Recently uploaded
Recently uploaded (20)
Inception Pack Vol 2: Bizarre premium
- 1. Cumulus: Filesystem Backup to the Cloud Michael Vrable, Stefan Savage, and Geoffrey M. Voelker Presented by Hakim Weatherspoon
- 2. Good „ole Days Replace your tape drives with something truly scalable
- 3. Amazon S3 to the rescue In Spring 2006, Amazon released a new storage API: Put, Get, List, Delete Build whatever you want! Quickly
- 4. Backing up the new way (S3) Smart Scales  - no longer our concern... Amazon's concern - all servers backup in parallel Cheap - old cost =XXXper year - new cost =YYYper year where YYY < XXX 
- 5. Thin vs Thick Cloud • E.g. Amazons S3 vs EMC‟s MozyPro • Thin – Can change provider easier – Applications can work across providers • Thick – Better performance – Locked into a provider – Provider can go out of business
- 6. Cumulus • Simple storage backup utility for Thin Clouds • Evaluates efficacy of cloud storage • Working prototype – http://www.cs.ucsd.edu/~mvrable/cumulus/
- 7. Outline • Motivation/Intro • Related Work • Design • Evaluation • Thoughts and Conclusions
- 8. Related Work
- 9. Outline • Motivation/Intro • Related Work • Design – API – Segments – Snapshots – Subfileincrementals – Cleaning – restoring • Evaluation • Thoughts and Conclusions
- 10. API • Same as S3 – Put, Get, List, Delete • Thin cloud – does not rely on integrated services – Can easily change provider and network protocols – S3, FTP, SFTP • WORM Model – Write-once, read-many – Requires writing new entirely file if changes occur – What are the cleaning overheads?
- 11. Segments • Aggregation via Segment Goals – Avoid costs due to small files • S3 charges on per file bases • Many small files – Avoid costs in network protocols • Small files have higher latency and other overheads – Compression • inter-file similarities – Privacy • Hide file boundaries • Negative consequences? – Need an entire segment to write
- 12. Snapshots
- 13. Sub-File Incrementals • Only stored changed part of files • New snapshots point to old objects when data unchanged • Byte ranges – portions of old objcts to be reused
- 14. Segment Cleaning • Similar to a log-structured file system (LFS) • Clean based on utilization of segment, α – α = 0, no cleaning – α = 1, clean with the slightest change • Cumulus – attempts to find an equilibrium for α – Uses a different process to clean – Marks a local database as “expired” – Then, next snapshot will not refer to expired segment
- 15. Restore • Full Restore – Download all segments for a snapshot • Partial Restore – Download snapshot descriptor, metadata, and only necessary segments • What happens if client machine dies? • How is latest snapshot descriptors identified? • What about sharing between client machines?
- 16. Outline • Motivation/Intro • Related Work • Design • Evaluation – Performance Case Study – Monetary Case Study • Thoughts and Conclusions
- 18. Backup over time (user trace)
- 19. Backup w/out Segment Cleaning (user trace)
- 20. Average Daily Storage (fileserver)
- 21. Average Daily Upload (fileserver)
- 22. Average Segments per Day (fileserver)
- 23. Storage overhead for 16MB Segment (fileserver)
- 25. Overheads
- 26. Monetary Case Study • Storage: $0.15 per GB . Month • Upload: $0.10 per GB • Segment: $0.01 per 1000 files uploaded • We are charged this amount, so please be careful with your labs and projects!!!
- 27. Monetary Costs for Backup
- 28. Costs for Backup (fileserver)
- 29. Monetary Cost Comparison (user trace)
- 30. Outline • Motivation/Intro • Related Work • Design • Evaluation • Thoughts and Conclusions
- 31. Discussion • Thoughts? • Did paper make case for Thin Clouds? • Sharing between clients ignored? • What every happened to P2P?!
- 32. Lab 0 • http://s3.amazonaws.com/edu-cornell-cs- cs5300/aws-get-started.html
- 33. Next Time • Read NFS and write review: – Design and Implementation of a Network File System, Sandberg, Goldberg, Kleiman, Walsh, and Lyon, USENIX 1985 • Do Lab 0 • Check website for updated schedule