Abstract Attribute-Based Prediction of File Properties(3)

The chi-squared test can also be used to rank the at-tributes by the degree of association.Figure 2shows how the chi-squared values differ for the size and lifespan properties.There are two important points to take from this ?gure.First,the attribute association differs across properties for a given trace –for example,in CAM-PUS the uid shows a relatively strong association with the lifespan,yet a weak association with the size.The second point is that the relative rankings differ across traces.For example,on CAMPUS the middle compo-nent of a ?le name has strong association with lifespan and size,but the association is much weaker on DEAS03and EECS03.

Although we show only two properties in these graphs,similarly diverse associations exist for other properties (e.g.,directory entry lifespan and read/write ratio).In Section 5we show how these associations can be dynamically discovered and used to make predictions.

The chi-squared test described in this section is a one-way test for association.This test provides statistical ev-idence that individual attributes are associated with ?le properties.It does not,however capture associations be-tween subsets of the attributes and ?le properties.It also does not provide an easy way to understand exactly what those associations are.One can extend this methodology to use -way chi-square tests,but the next section dis-cusses a more ef?cient way for both capturing multi-way associations and extracting those associations ef?ciently.

5

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.

5The ABLE Predictor

The results of the previous section establish that each of a?le’s attributes(?le name,uid,gid,mode)are,to some extent,associated with its long term properties (size,lifespan,and access pattern).This fact suggests that these associations can be used to make predictions on the properties of a?le at creation time.The chi-squared results also give us hope that higher order as-sociations(i.e.,an association between more than one at-tribute and a property)may exist,which could result in more accurate predictions.

To investigate the possibility of creating a predictive model from our data,we constructed an Attribute-Based Learning Environment(ABLE).ABLE is a learning en-vironment for evaluating the predictive power of?le at-tributes.The input to ABLE is a table of information about?les whose attributes and properties we have al-ready observed and a list of properties for which we wish to predict.The output is a statistical analysis of the sam-ple,a chi-squared ranking of each?le attribute relative to each property,and a collection of predictive models that can be used to make predictions about new?les.

In this paper,we focus on three properties:the?le size,the?le access pattern(read-only or write-only),and the?le lifespan.On UNIX?le systems,there are two as-pects of?le lifespan that are interesting:the?rst is how long the underlying?le container(usually implemented as an inode)will live,and the other is how long a par-ticular name of a?le will live(because each?le may be linked from more than one name).We treat these cases separately and make predictions for each.

To simplify our evaluation,each property we wish to predict is represented by a Boolean predicate.For exam-ple:

size

size16KB

inode lifespan1sec

?le name lifespan1sec

read-only

write-only

We believe these properties are representative of properties that a?le or storage system designer might use to optimize for different classes of?les.For exam-ple,if we know that a?le will be read-only,then we might choose to replicate it for performance and avail-ability,but this optimization would be inappropriate for ?les that are written frequently but rarely read.Write-only?les might be stored in a partition optimized for writes(e.g.,a log-structured?le system),and short-lived ?les could live their brief lives in NVRAM.In Section

6,for example,we show that by identifying small,short-lived?les and hot directories,we can use predictions to optimize directory updates in a real?le system.

ABLE consists of three steps:

Step1:Obtaining Training Data.Obtain a sample of ?les and for each?le record its attributes(name,

uid,gid,mode)and properties(size,lifespan,and

access pattern).

Step2:Constructing a Predictive Classi?er.For each ?le property,we train a learning algorithm to clas-

sify each?le in the training data according to that

property.The result of this step is a set of predic-

tive models that classi?es each?le in the training

data and can be used to make predictions on newly

created?les.

Step3:Validating the e the model to pre-dict the properties of new?les,and then check

whether the predictions are accurate.

Each of these steps contains a number of interesting issues.For the?rst step,we must decide how to obtain representative samples.For the second,we must choose

a learning algorithm.For the third,we must choose how

to evaluate the success of the predictions.We may con-sider different types of errors to have different degrees of importance–for example,if the?le system treats short-lived?les in a special manner,then incorrectly predicting that a?le will be short-lived may be worse than incor-rectly predicting that a?le will be long-lived.

5.1Obtaining Training Data

There are two basic ways to obtain a sample of?les: from a running system or from traces.ABLE currently uses the latter approach,using the NFS traces described in Section3.

Determining some of the attributes of a?le(gid,uid, mode)is a simple matter of scanning the traces and cap-turing any command(e.g.,lookup or getattr) …… 此处隐藏：6170字，全部文档内容请下载后查看。喜欢就下载吧 ……