uStorage - A storage architecture to provide Block-level Storage through Object-based Storage
Block-level Storage is widely used to support heavy work- loads. It can be directly accessed by the operating system, but it faces some durability issues, hardware limitations and performance degradation in geographically distributed systems. Object-based Storage Device (OSD) is a data storage concept widely used to support write-once-read-many (WORM) systems. Because OSD contains data, metadata and an unique identifier, it becomes very powerful and customizable. OSDs are ideal for solving the increasing problems of data growth and resilience requirements while mitigating costs. This paper describes a scalable storage architecture that uses OSD from a distributed P2P Cloud Storage system and delivers a Block-level Storage layer to the user. This architecture combines the advantages of the replication, reliability, and scalability of a OSD on commodity hardware with the simplicity of raw block for data-intensive workload. We retrieve data from the OSD in a set of blocks called buckets, allowing read-ahead operations to improve the performance of the raw block layer. Through this architecture we show the possibility of using OSD on the back end and deliver a storage layer based on raw blocks with better performance to the end user. We evaluated the proposed architecture based on the cache behavior to understand non-functional properties. Experiments were performed with different cache sizes. High throughput performance was measured for heavy workloads at the two storage layers.
Recommended
Recommended
More Related Content
What's hot
What's hot (17)
Similar to uStorage - A storage architecture to provide Block-level Storage through Object-based Storage
Similar to uStorage - A storage architecture to provide Block-level Storage through Object-based Storage (20)
Recently uploaded
Recently uploaded (20)
uStorage - A storage architecture to provide Block-level Storage through Object-based Storage
- 1. uStorage - A storage architecture to provide Block-level Storage through Object-based Storage Felipe Oliveira Gutierrez felipe.o.gutierrez@gmail.com UFPE - Federal University of Pernambuco - Brazil Ustore Cloud Services September 28, 2017 Vinicius C. Garcia, José Fernando Cardoso, Thiago Jamir, Josino Rodrigues Neto, Rodrigo Assad, and Marcos Barreto
- 2. Agenda ● Motivation ● Problem ● State of the art ● Architecture ○ proposal ○ implementation ○ evaluation ● Conclusion and future work 2 ESOCC 2017Sep 28th, 2017
- 3. 3 ESOCC 2017Sep 28th, 2017
- 4. Block Storage ● raw block ● same size ● no metadata ● scalability limited ● ID to local system ● good performance OS mount point 4 Object Storage ● buckets ● different size ● expandable metadata ● very scalable ● global unique ID on distributed systems ● bad performance OS mount point ESOCC 2017Sep 28th, 2017
- 5. Block Storage ● needs HA for prepare to fail ● high performance hardware ● reduced availability ● random RW loads ● structured data sets 5 Object Storage ● already prepared to fail ● hardware flexibility ● high availability ● generally read ● unstructured data sets ESOCC 2017Sep 28th, 2017
- 6. Other architectures that rely on Block Storage model 6 NexentaStor Ceph Panasas File-based Storage, Block-level Storage File-based Storage, Block-level Storage, Object-based Storage Extended Object File System (exofs) Serviço de Cloud Storage POSIX filesystem RAID on the cluster Filesystem replication on different domains Fault tolerance through data replication Panasas Active Scale Storage Cluster ESOCC 2017Sep 28th, 2017
- 7. uStorage architecture 7 ESOCC 2017Sep 28th, 2017
- 8. uStorage architecture 8 ESOCC 2017Sep 28th, 2017
- 9. uStorage architecture 9 ESOCC 2017Sep 28th, 2017
- 10. uStorage architecture 10 ESOCC 2017Sep 28th, 2017
- 11. uStorage architecture 11 ESOCC 2017Sep 28th, 2017
- 12. uStorage architecture implementation 12 ESOCC 2017Sep 28th, 2017
- 13. Raw block vs Bucket 13 ESOCC 2017Sep 28th, 2017
- 14. uStorage architecture requirements ● Which clean policy must be used on the iSCSI Target cache? ● When the clean policy must be executed? ● How many buckets must be removed on the policy execution? ● Which buckets must be removed from the iSCSI Target cache? 14 ESOCC 2017Sep 28th, 2017
- 15. 15 Sequence diagram of the cache algorithm (iSCSI Target) ESOCC 2017Sep 28th, 2017
- 16. Finite State Machine (FSM) of the buckets on the cache algorithm (iSCSI Target) 16 ESOCC 2017Sep 28th, 2017
- 17. Trade-off of the bucket size (2 MiB) 17 ESOCC 2017Sep 28th, 2017
- 18. 18 Cache LRU algorithm ESOCC 2017Sep 28th, 2017
- 19. 19 Cache LRU algorithm ESOCC 2017Sep 28th, 2017
- 20. 20 Cache LRU algorithm ESOCC 2017Sep 28th, 2017
- 21. 21 Cache LRU algorithm ESOCC 2017Sep 28th, 2017
- 22. 22 Cache LRU algorithm ESOCC 2017Sep 28th, 2017
- 23. 23 Cache LRU algorithm ESOCC 2017Sep 28th, 2017
- 24. 24 Cache LRU algorithm ESOCC 2017Sep 28th, 2017
- 25. Cache size: 20 GiB - Files: 20 GiB 25 ESOCC 2017Sep 28th, 2017
- 26. Cache size: 5 GiB - Files: 20 GiB 26 ESOCC 2017Sep 28th, 2017
- 28. Multiple LUNs on the same iSCSI Target 28 ESOCC 2017Sep 28th, 2017
- 29. Conclusion and future work ● The main configuration that improves the performance is the bucket size. ○ Better evaluation of the bucket size. ● Optimize the restore operations if we want more IOPs on the cache. ● Switch the relational database on the Metadata Storage for a faster database. ● Improve cache readahead through bucket prefetching. ○ The system init sequence doesn’t change. ○ Watch file read patterns. ○ Improve disk throughput. 29 ESOCC 2017Sep 28th, 2017
- 30. uStorage - A storage architecture to provide Block-level Storage through Object-based Storage Felipe Oliveira Gutierrez felipe.o.gutierrez@gmail.com UFPE - Federal University of Pernambuco - Brazil Ustore Cloud Services September 28, 2017 Vinicius C. Garcia, José Fernando Cardoso, Thiago Jamir, Josino Rodrigues Neto, Rodrigo Assad, and Marcos Barreto