Abstract Attribute-Based Prediction of File Properties

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.

Attribute-Based Prediction of File Properties Daniel Ellard,Michael Mesnier,Eno Thereska,Gregory R.Ganger,Margo Seltzer Abstract

We present evidence that attributes that are known to

the?le system when a?le is created,such as its name,

permission mode,and owner,are often strongly related

to future properties of the?le such as its ultimate size,

lifespan,and access pattern.More importantly,we show

that we can exploit these relationships to automatically

generate predictive models for these properties,and that

these predictions are suf?ciently accurate to enable opti-

mizations.

1Introduction

In“Hints for Computer System Design,”Lampson

tells us to“Use hints to speed up normal execution.”[14]

The?le system community has rediscovered this prin-ciple a number of times,suggesting that hints about a ?le’s access pattern,size,and lifespan can aid in a va-riety of ways including improving the?le’s layout on disk and increasing the effectiveness of prefetching and caching.Unfortunately,earlier hint-based schemes have required the application designer or programmer to sup-ply explicit hints using a process that is both tedious and error-prone,or to use a special compiler that can recog-nize speci?c I/O patterns and automatically insert hints. Neither of these schemes have been widely adopted.

In this paper,we show that applications already give useful hints to the?le system,in the form of?le names and other attributes,and that the?le system can success-fully predict many?le properties from these hints.

We begin by presenting statistical evidence from three contemporary NFS traces that many?le attributes,such as the?le name,user,group,and mode,are strongly re-lated to?le properties including?le size,lifespan,and access patterns.We then present a method for automati-cally constructing tree-based predictors for the properties of a?le based on these attributes and show that these

predictions are accurate.Finally,we discuss uses for such predictions,including an implementation of a sys-tem that uses them to improve?le layout by anticipating which blocks will be the most frequently accessed and grouping these blocks in a small area on the disk,thereby improving reference locality.

The rest of this paper is organized as follows:Sec-tion2discusses related work.Section3describes the collection of NFS traces we analyze in this study.Sec-tion4makes the case for attribute-based predictions by presenting a statistical analysis of the relationship be-tween attributes of?les and their properties.Section5 presents ABLE,a classi?cation-tree-based predictor for several?le properties based on their attributes.Section6 discusses how such models might be used,and demon-strates an example application which increases the local-ity of reference for on-disk block layout.Section7con-cludes.

2Related Work

As the gap between I/O and CPU performance has increased many efforts have attempted to address it.An entire industry and research community has emerged to 1

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.

attack I/O performance;?le systems have been modi?ed, rewritten and rethought in attempts to reduce the number of synchronous disk requests.Signi?cant effort has also been expended to make caches more effective so that the number of disk requests can be reduced.Many powerful heuristics have been discovered,often from the analyses of real workloads,and incorporated into production?le systems.All of these endeavors have been productive, but I/O performance is still losing ground to CPU,mem-ory,and network performance,and we have not resolved the I/O crisis to which Patterson refers in the original RAID paper,written more than?fteen years ago[24].

There is extensive ongoing research in the?le system and database communities regarding the optimization of various aspects of performance,reliability,and availabil-ity of data access.Many heuristics have been developed and incorporated into popular?le systems like the Fast File System(FFS)[17].Many of these heuristics depend on assumptions about workloads and?le properties.

One example of a contemporary?le system is the Fast File System(FFS)[17],whose basic design is nearly twenty years old and yet continues to be tuned[7].For example,FFS is optimized to handle small?les in a dif-ferent manner than large?les;it attempts to organize small?les on disk so that they are near their metadata and other?les in the directory,under the assumption that ?les in the same directory are often accessed together. Some?le systems go to more extreme lengths,such as storing the contents of short?les in the same disk block as their inode[22]or storing the directory and inode in-formation in the same block[11].

In addition to size,other properties of?les,such as whether they are write-mostly or read-mostly,have been found useful to drive various?le system policies.For example,the assumption underlying the design of the log-structured?le system(LFS)is that write-latency is the bottleneck for?le system performance[26].Hybrid schemes that use LFS to store write-mostly?les have also found this approach useful[23].In contrast,if a ?le is known to be read-mostly,it may bene?t from ag-gressive replication for increased performance and avail-ability[27].

Unfortunately,every widespread heuristic approach suffers from at least one of the following problems:First, if the heuri …… 此处隐藏：6030字，全部文档内容请下载后查看。喜欢就下载吧 ……