Energy Efficiency and Storage Flexibility in the Blue File System

1
Energy Efficiency and Storage Flexibility in the
Blue File System

• Edmund B Nightingale
• Jason Flinn
• University of Michigan

2
Towards Ubiquitous Data Access

• Pervasive computing anywhere anytime any device
• Ubiquitous WiFi access
• Mobile storage
• Barriers to pervasive data access
• Varying I/O performance
• Energy hogs!
• Stale, inconsistent data
• BlueFS overcomes these barriers

3
Change is Constant

• Access times vary across storage devices
• Wireless networks are highly variable
• Power management affects performance
• BlueFS provides an adaptive cache hierarchy

4
Power to the People

• Distributed file systems use power hungry devices
• Distributed file systems not designed for energy
• BlueFS clean sheet design for energy efficiency

5
You Can Take it with You

• Want to leverage the benefits of portable storage
• Portable storage is not a mobile data panacea
• Can be lost, stolen, damaged
• Sharing is hard
• More replication higher chances of
  stale/inconsistent data
• BlueFS integrates portable storage into file
  system

6
Organization

• Motivation
• The Blue File System
• Evaluation
• Related Work
• Conclusion

7
The Blue File System
Work
Home
BlueFS Server
gt /BlueFS/
8
Write to Many
VFS Ops

• Asynchrony improves performance
• Aggregating writes saves energy

To microdrive
To BlueFS Server
BlueFS
9
Read From Any
gt cat /BlueFS/foo
Blk 1
BlueFS Server
BlueFS
Blk 2

• BlueFS monitors cost of reading data from each
  device
• Pick device best performance, least energy

10
Power Management
Hints
State Transitions
BlueFS

• BlueFS integrated with power management
  strategies
• When a device transitions, BlueFS is notified

11
Hiding device transitions
5) Network transitions to PSM
1) Network accesses
Network
2) Ghost hints issued
3) Disk spins up
4) Disk accesses begin
Disk
12
Cache Management

• Storage devices treated as cache
• BlueFS decides which files to cache
• When a device fills up, LRU used to make space
• Affinity file always on portable storage

13
Cache Consistency
Work
Home
BlueFS Server
gt /BlueFS/
gt /BlueFS/
14
Evaluation Goals

• How does performance of BlueFS compare to other
  DFS?
• What benefit does an adaptive cache hierarchy
  provide?
• How well does BlueFS support portable storage?
• How much does BlueFS reduce client energy usage?
• How does BlueFS perform with heterogeneous
  storage?

15
Comparing BlueFS to other DFS

• Andrew benchmark Untar, configure, make, delete
  Apache source

6000.0

• BlueFS 10x faster than NFS, 16 faster than Coda
  WD

16
Benefits of Adaptive Cache Hierarchy

• Read various files with the disk in standby mode

• BlueFS reads 4k files up to 60x faster than ext2

17
Support for Portable Storage

• Ran a file system trace as fast as possible using
  portable storage

• BlueFS 44 faster than Lookaside, up to 3x faster
  for high latency

18
Energy Efficient Data Access

• Ran first 10,000 ops of Purcell trace with half
  full cache

• Reduces interactive delay 76, 55 less energy
  than Coda

19
Heterogeneous Storage

• Added 16 MB flash to BlueFS, ran Purcell with
  full local cache

• With flash BlueFS reduces interactive delay 48,
  25 less energy

20
Related Work

• AFS Howard et al
• Coda Kistler et al lookaside caching Tolia
  et al
• PersonalRAID Sobti et al
• Segank Sobti et al
• Cooperative I/O Weissel et al
• LBFS Muthitacharoen et al

21
Conclusions

• BlueFS enables ubiquitous data access by
• Adapting to device characteristics
• Reducing mobile computer energy usage
• Seamlessly integrating portable storage
• Source http//notrump.eecs.umich.edu/group/group.
  html

22
System Architecture

23
BFS Kernel

• Intercepts Linux VFS calls
• Communicates with Wolverine through pseudo-device
  driver
• Caches and invalidates data metadata in Linux
  file cache
• Operations redirected to Wolverine
• Enables support for devices with different
  consistency semantics