Policies in Distributed Data Storage

1
Policies in Distributed Data Storage

• Presented by Ajay Potnis
• Professor Craig E. Wills
• December 2000

2
Acknowledgements

• Professor Craig E. Wills
• Uresh Vahalia
• Xiaoye Jiang
• Gary Ma
• Jiannan Zheng
• Yael Melman
• Mrunal Potnis
• Aditya Potnis

3
Goal

• Study various distributed architectures
• NFS
• xFS
• SAN
• SSFS
• Conclusion

4
Direct Attached Storage
Data Storage Unit
Data on SCCI/Fibre
5
Direct Attached Storage

• Channel I/O Speed
• Can select the desired storage
• No sharing
• No flexibility

6
NFS
Data Storage Unit
Data and Metadata
Data on SCCI/Fibre
File Server
7
NFS

• Simple client server semantics using LAN/WAN
• Server is the owner of file system
• Uses RPC for communication
• NFS with Mount protocol makes FS locally
  available
• Support for advisory locking using NLM
• Caching in memory/on disk improves performance on
  clients (stale data)
• Data can be shared by multiple clients
• Administration is easy due to centralization of
  data
• Server can provide scalability, high
  availability, online maintenance, etc.

8
xFS
Client
Client and Manager
Forward Request
Data Storage Unit
Data
Request
Forward Request
Data on SCCI/Fibre
Data
Storage Server
9
xFS

• Server-less architecture (clients, storage
  servers and managers)
• Distributes cache, storage and metadata
  management over machines
• Uses a stateful protocol
• Uses Active Messages instead of RPC for
  communication
• Multiple managers could lead to more network
  traffic
• Flexibility adding many clients, managers and
  storage servers
• Client write data to storage server before giving
  it to another client to avoid consistency
  problems in failures
• Increased circular dependencies and complexities

10
SAN (switched fabric)
File Server
Data Storage Unit
Switch
Data
Data
Data on Fibre
11
SAN

• Connects multiple clients to multiple storage
  drives
• Storage can be configured at will
• Needs good management tools and good Sys Admin
  skills
• No seamless sharing of same storage by multiple
  clients/servers

12
SSFS
Client 1
Data on SCCI / Fibre
1.Metadata Request
2. Metadata
Data Storage Unit uses SAN
Data on SCCI / Fibre for NAS
Server
2. Metadata
1.Metadata Request
Data on SCCI / Fibre
Client 2
13
SSFS

• Client-Server-Storage semantics using NAS and SAN
• NAS for sharing, SAN for channel I/O speed
• File sharing among heterogeneous clients
• Concurrent sharing among NAS and SAN clients
• Server is the owner of metadata and storage
• RPC communication for metadata exchange over
  network
• Caching in memory improves performance on clients
  (stale data not possible because of callbacks)
• Read/Write throughput increased since data goes
  at channel speed (can use multiple SCSI/fibre
  paths to write)
• Multiple clients can use the same storage device
• Scalability, high availability, transparent to
  applications
• Failure recovery built in the protocol

14
Conclusion

• Why not get two birds in one shot?