Integrating Portable and Distributed Storage · Integrating Portable and Distributed Storage Niraj...
Transcript of Integrating Portable and Distributed Storage · Integrating Portable and Distributed Storage Niraj...
Integrating Portable and Distributed Storage
Niraj Tolia
Jan Harkes, Michael Kozuch, M. Satyanarayanan
PARALLEL DATA LABORATORYINTEL RESEARCH PITTSBURGH
Carnegie Mellon University
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 2
Motivation
• Explosive growth of portable storage devices
• Are these glorified floppy disks?
• In a world of ever-improving networking, why do we still carry them around?
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 3
Integration Value
• Portable Devices• Performance• Availability• Ubiquity
• Distributed File Systems• Robustness• Sharing and Collaboration• Consistency• Capacity
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 4
10,000 Foot View
• Lookaside Caching - All cache lookups check portable device
• Minimal disruption of existing usage model
Distributed File System Server
Client
NetworkNetwork
PortableDevice
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 5
Outline
• Motivation• Lookaside Caching• Benchmarks and Evaluation• Related Work• Conclusion
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 6
Lookaside Caching
• Technique to combine the strengths of dist. file systems and portable storage
• Design Goals• Be tolerant of user error• No compromise of robustness or consistency
• Simplicity!
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 7
Design Decisions
• File server is the authoritative source• No change made to native file system
• Allows flexibility in the choice of device• Allows user to look at files stored on the
portable device without a file system client• Only addition is an index file
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 8
Lookaside Caching Flowchartopen()
cached
done network fetch
done
Yes No
fetch attributes
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 9
Lookaside Caching Flowchartopen()
cached
done network fetch
done
Yes No
fetch attributes + hash
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 10
Lookaside Caching Flowchartopen()
cached
done
Yes No
fetch attributes + hash
portable
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 11
Lookaside Caching Flowchartopen()
cached
done
Yes No
fetch attributes + hash
portable
done
Yes
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 12
Lookaside Caching Flowchartopen()
cached
done
network fetch
done
Yes No
fetch attributes + hash
portable
done
Yes No
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 13
Implementation• File system metadata extended to
include a cryptographic hash (SHA-1) • Based upon whole-file hashing• Implemented in Coda 6.0
• mkdb utility to generate index file• Index generated from normal file system tree• Lazy update process• Allows users to change data
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 14
Outline
• Motivation• Lookaside Caching• Benchmarks and Evaluation
• Benchmark Descriptions• Evaluation Setup• Results
• Related Work• Conclusion
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 15
Benchmarks
• Kernel Compile • Client File System Trace • Virtual Machine Migration
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 16
Benchmark Description
• Kernel Compile• Compiled Linux 2.4.18• Compile reads 80.7 MB (118 MB source tree)• Portable device contained either 2.4.18,
2.4.17, 2.4.13, 2.4.9, or 2.4.0
• Client File System Traces• Four desktop clients from the DFSTrace traces• No think time modelled
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 17
Benchmark Description
• Virtual Machine Migration• Based on the Internet Suspend/Resume project• Time taken to resume a migrated Virtual
Machine and execute a MS Office-based productivity benchmark
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 18
Evaluation Setup
• Hi-Speed USB Flash device• Evaluated at different bandwidths
• 100 Mb/s, 10 Mb/s, 1 Mb/s + 10 ms, and 100 Kb/s + 100 ms
NIST Net Router
File ServerClient
USB Keychain
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 19
USB Portable Device Bandwidth
• Max below manufacturer ratings• Low bandwidth observed for small files
25.016.764 KB
26.525.8
7.4Write (Mb/s)
29.4100 MB28.61 MB
6.34 KBRead (Mb/s)Size (KB)
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 20
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
100 Mb/s 10 Mb/s 1 Mb/s 100 Kb/s
No
rmal
ized
Ru
nti
me
2.4.0 .9 .13 .17 .18
Linux Kernel Compile
Linux Kernel Versions
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 21
Trace Replay
0.0
0.2
0.4
0.6
0.8
1.0
1.210
0 M
b/s
10 M
b/s
1 M
b/s
100
Kb/
s
100
Mb/
s
10 M
b/s
1 M
b/s
100
Kb/
s
Berlioz Purcell
No
rmal
ized
Ru
nti
me
33% 66% 100%
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 22
Internet Suspend/Resume
• Two metrics of performance• Resume Time• Benchmark Time
VM started VM usable
Time
End of Benchmark
Resume Time
BenchmarkTime
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 23
Internet Suspend/Resume
• Resume Time• Portable device had VM memory image (100%
hit rate)
99.7%124301100 Kb/s
96.2%123171 Mb/s
69.2%123910 Mb/s
7.1%1314100 Mb/s
Win %With Lookaside (s)No Lookaside (s)
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 24
Internet Suspend/Resume
• Benchmark Time• Portable device had an unmodified Windows
XP + Office image (~50% hit rate)
• Slowdown
68.8%953030531100 Kb/s
61.6%103226881 Mb/s
42.7%21237010 Mb/s
6.9%161173100 Mb/s
Win %With Lookaside (s)No Lookaside (s)
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 25
Internet Suspend/Resume
• Suspend Time• Lookaside caching does not currently handle this• Apart from portable storage, you can also use
staging servers, etc.• Topic of ongoing research
VM started VM usable
Time
End of Benchmark
Resume Time
BenchmarkTime
Data Safe
SuspendTime
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 26
Related Work
• Mobile Storage Solutions• Segank, PersonalRaid, Footloose, Personal
Server
• Commercial Offerings• Migo, KeyComputing
• Other Content Addressable Systems• CASPER, LBFS• CFS/Chord, PAST/Pastry, Venti, Pastiche
Niraj Tolia April 2, 2004http://www.pdl.cmu.edu/ 27
Conclusion
• Demonstrated the integration of portable and distributed storage
• Shown that a simple optimization can have a dramatic effect on performance
• Future Work• While invented for portable devices, lookaside
caching can be extended to use off-machine storage