Abstract Attribute-Based Prediction of File Properties(6)

We have shown that the attributes of a?le are strong hints of how that?le will be used.Furthermore,we have exploited these hints to make accurate predictions about the longer-term properties of?les,including the size, read/write ratio,and lifespan.Overall,?le names pro-vide the strongest hints,but using additional attributes can improve prediction accuracy.In some cases,accu-rate predictions are possible without considering names at ing traces from three NFS environments,we have demonstrated how classi?cation trees can predict ?le and directory properties,and that these predictions can be used within an existing?le system.

Our results are encouraging.Contemporary?le sys-tems use hard-coded policies and heuristics based on general assumptions about their workloads.Even the most advanced?le systems do no more than adapt to vio-lations of these assumptions.We have demonstrated how to construct a learning environment that can discover pat-terns in the workload and predict the properties of new ?les.These predictions enable optimization through dy-namic policy selection–instead of reacting to the prop-erties of new?les,the?le system can anticipate these properties.Although we only provide one example?le system optimization(clustering of hot directory data), this proof-of-concept demonstrates the potential for the system-wide deployment of predictive models.

ABLE is a?rst step towards a self-tuning?le system or storage device.Future work involves automation of the entire ABLE process,including sample collection, attribute selection,and model building.Furthermore, since changes in the workload will cause the accuracy of our models to degrade over time,we plan to auto-mate the process of detecting when models are failing (or are simply suboptimal)and retraining.When cata-clysmic changes in the workload occur(e.g.,tax season in an accounting?rm,or September on a college cam-pus),we must learn to detect that such an event has oc-curred and switch to a new(or cached)set of models. We also plan to explore mechanisms to include the cost of different types of mispredictions in our training in or-der to minimize the anticipated total cost of errors,rather than simply trying to minimize the number of errors.

In addition to caching and on-disk layout optimiza-tion,we envision a much larger class of applications that will bene?t from dynamic policy selection.Attribute-based classi?cation of system failures and break-ins(or anomaly detection)is a natural adjunct to this work

(e.g.,“has this?le been compromised?”).Moreover,

through the same clustering techniques implemented by our decision trees,we feel that semantic clustering can be useful for locating information(e.g.,“are these?les related?”).Both of these are areas of future work.

Acknowledgments

Daniel Ellard and Margo Seltzer were sponsored in part by IBM.The CMU researchers thank the mem-bers and companies of the PDL Consortium(including EMC,Hewlett-Packard,Hitachi,IBM,Intel,Microsoft, Network Appliance,Oracle,Panasas,Seagate,Sun,and Veritas)for their interest,insights,feedback,and sup-port.Their work is partially funded by the National Sci-ence Foundation,via grants#CCR-0326453and#CCR-0113660.

References

[1]Sedat Akyurek and Kenneth Salem.Adap-

tive Block puter Systems,

13(2):89–121,1995.

[2]J.Michael Bennett,Michael A.Bauer,and David

Kinchlea.Characteristics of Files in NFS Environ-

ments.In Proceedings of ACM SIGSMALL Sympo-

sium on Small Systems/PC.,pages18–25,Toronto,

Ontario,Canada,1991.

[3]Matthew A.Blaze.NFS Tracing by Passive Net-

work Monitoring.In Proceedings of the USENIX

Winter1992Technical Conference,pages333–343,

San Fransisco,CA,January1992.

[4]Leo Breiman,Jerome H.Friedman,Richard A.Ol-

shen,and Charles J.Stone.Classi?cation and Re-

gression Trees.Chapman and Hall,1984.

[5]Pei Cao,Edward W.Felten,Anna R.Karlin,

and Kai Li.Implementation and Performance

of Integrated Application-Controlled File Caching,

Prefetching,and Disk Scheduling.ACM Transac-

tions on Computer Systems,14(4):311–343,1996.

[6]Rich Caruana and Dayne Freitag.Greedy Attribute

Selection.In International Conference on Machine

Learning,pages28–36,1994.

[7]Ian Dowse and David Malone.Recent Filesystem

Optimisations on FreeBSD.In Proceedings of the 13

We present evidence that attributes that are known to the file system when a file is created, such as its name, permission mode, and owner, are often strongly related to future properties of the file such as its ultimate size, lifespan, and access pattern.

USENIX Annual Technical Conference(FREENIX Track),pages245–258,June2002.

[8]Daniel Ellard,Jonathan Ledlie,Pia Malkani,and

Margo Seltzer.Passive NFS Tracing of Email and Research Workloads.In Proceedings of the Second USENIX Conference on File and Storage Technologies(F AST’03),pages203–216,San Fran-cisco,CA,March2003.

[9]Daniel Ellard,Jonathan Ledlie,and Margo Seltzer.

The Utility of File Names.Technical Report TR-05-03,Harvard University Division of Engineering and Applied Sciences,2003.

[10]Daniel Ellard and Margo Seltzer.New NFS Tracing

Tools and Techniques for System Analysis.In Pro-ceedings of the Seventeenth Annual Large Installa-tion System Administration Conference(LISA’03), pages73–85,San Diego,CA,October2003. [11]Gregory R.Ganger and M.Frans Kaashoek.Em-

bedded Inodes and Explicit Grouping:Exploiting Disk Bandwidth for Small Files.In USENIX An-nual Technical Conference,pages1–17,1997. [12]Gregory R.Ganger,Marshall Kirk McKusick,

Craig A.N.Soules,and Yale N.Patt.Soft Updates:

a Solution to the Metadata Update Problem in File

Systems.ACM Transactions on Computer Systems, 18(2):127–153,2000.

[13]Trevor Hastie,Robert Tibshirani,and Jerome

Friedman.The Elemen …… 此处隐藏：5653字，全部文档内容请下载后查看。喜欢就下载吧 ……