A PREDICTIVE AUTOMATIC TUNING SERVICE FOR OBJECT

POOLING BASED ON DYNAMIC MARKOV MODELING

Nima Shariﬁmehr and Samira Sadaoui

Department of Computer Science, University Of Regina, Regina, Canada

Keywords:

Object pool service, markov model, prediction, automatic tuning, workload modeling, enterprise applications.

Abstract:

One of the most challenging concerns in the development of enterprise software systems is how to manage

effectively and efﬁciently available resources. Object pooling service as a resource management facility signif-

icantly improves the performance of application servers. However, tuning object pool services is a complicated

task that we address here through a predictive automatic approach. Based on dynamic markov models, which

capture high-order temporal dependencies and locally optimize the required length of memory, we ﬁnd pat-

terns across object invocations that can be used for prediction purposes. Subsequently, we propose an effective

automatic tuning solution, with reasonable time costs, which takes advantage of past and future information

about activities of object pool services. Afterwards, we present experimental results which demonstrate the

scalability and effectiveness of our novel tuning solution, namely predictive automatic tuning service.

1 INTRODUCTION

One of the most challenging concerns in the develop-

ment of enterprise software systems is how to manage

effectively and efﬁciently available resources and re-

alize user requirements. Resource management must

be applied across both low-level resources (i.e., CPU

time, memory usage) and high-level resources (i.e.,

database connections) (Jordan et al., 2004). Con-

sequently, to improve the performance of resource

management, applicable techniques, such as resource

pooling (Kircher and Jain, 2004) and distributed re-

source allocation (Raman et al., 2003) have been in-

troduced. Here we are interested in resource pooling

which not only improves scalability through sharing

the cost of resource initialization but also allows accu-

rate tuning of memory usage (Crawford and Kaplan,

2003). Enterprise application servers (Oberle et al.,

2004) as sophisticated middlewares signiﬁcantly take

advantage of this technique to accelerate access to

resources such as persistent resources, database con-

nections and execution threads. In particular, object-

oriented application servers provide object pool ser-

vices as a special case of resource pooling. For ex-

ample, IBM WebSphere as a J2EE application server

prepares object pool service for entity beans (Ander-

son and Anderson, 2002).

However, object pool service tuning is a compli-

cated task which has not been addressed precisely.

Each object pool service contains a set of object pools

whose sizes should be tuned efﬁciently. Manual tun-

ing of object pools is extremely diffcult, and because

of its static nature it can not adapt itself with the

changing characteristics of workloads over time (Sul-

livan et al., 2004). In contrast, automatic tuning can

be developed in a dynamic way that is adaptive con-

sidering the variable nature of workloads. A tradi-

tional automatic tuning solution works only accord-

ing to past information (Barrera, 1993) and ignores

the future information that can be extracted from the

workload model.

In this paper, we propose a novel solution, namely

Predictive Automatic Tuning Service (PATS), for ob-

ject pooling. PATS takes into consideration past

and future information. In other words, we show

that tuning through prediction of object behaviors

achieves higher performance comparing to traditional

approaches. Different types of Markov models have

been used to study stochastic processes and it is

shown that they are efﬁcient tools for modeling and

Shariﬁmehr N. and Sadaoui S. (2007).

A PREDICTIVE AUTOMATIC TUNING SERVICE FOR OBJECT POOLING BASED ON DYNAMIC MARKOV MODELING.

In Proceedings of the Second International Conference on Software and Data Technologies - Volume ISDM/WsEHST/DC, pages 38-45

DOI: 10.5220/0001329700380045

 SciTePress

Figure 1: Integration of PATS into the middleware-based enterprise application architecture. This proposed integration

approach lets PATS monitor requests sent from each client to different Session Facade Components (SFCs). Monitoring

incoming requests is a prerequisite to extract clients behavioral patterns. On the other hand, placing PATS inside the business

logic layer is a necessity to be able to send tuning signals to the object pooling service.

predicting users’ behavior (Deshpande and Karypis,

2004). Thus, here for prediction purposes, we use

Dynamic Markov Modeling (DMM) (Cormack and

Horspool, 1987). DMM, which has been introduced

in Dynamic Markov Compression (DMC) method

(Cormack and Horspool, 1987), is a variant of Vari-

able Length Markov Model (VLMM) (Stefanov et al.,

2005). Thanks to DMM, we model the sequence of

callings on different objects which use a given object

pool service. Through this modeling process, we ﬁnd

patterns across object invocations that can be used as a

prediction tool. PATS is fully implemented with Java

(a total of 2,000 lines of code). In this paper, we con-

duct several experimentations to demonstrate the ef-

fectiveness of PATS in practice.

The rest of this paper is structured as follows.

Section 2 describes how PATS can be integrated into

enterprise applications. Section 3 introduces our be-

havior prediction approach based on DMM. Section

4 explores the detailed design of our automatic tuning

approach. Section 5 illustrates the results of workload

simulation that evaluates the performance of PATS.

Finally, Section 6 concludes the paper with some fu-

ture directions.

2 PATS INTEGRATION INTO

ENTERPRISE APPLICATIONS

The integration of PATS into enterprise applications is

an important issue that we address here (cf. Figure 1 ).

Most current enterprise applications are composed of

four logical layers (Ebner et al., 2000): client-side

and server-side presentation layers, business logic

layer and data layer. In middleware-based enterprise

applications, the business logic layer is generally im-

plemented inside an application server. This layer is

composed of many business components which inter-

act with object pools. However, many business pro-

cesses involve complex manipulation of these busi-

ness components. Therefore, Session Facade Pat-

tern (Alur et al., 2001) should be used as a higher-

level business abstraction that encapsulates interac-

tions among lower-level business components. In

other words, clients must only access application fea-

tures through Session Facade Components (SFCs)

which implement Session Facade Pattern.

According to our integration approach, object

pooling services must assign a speciﬁc part of the

object pool, namely Object Pooling Region (OPR),

with a speciﬁc size and replacement policy to each

SFC. So, objects which are being manipulated by

A PREDICTIVE AUTOMATIC TUNING SERVICE FOR OBJECT POOLING BASED ON DYNAMIC MARKOV

MODELING

each SFC can use assigned regions for object pool-

ing purposes. The determination of replacement pol-

icy for each OPR has been already successfully ad-

dressed by others (Guo and Solihin, 2006). In this pa-

per, we focus exclusively on tuning the size of each

OPR. This is a necessary task that signiﬁcantly af-

fects the applicability of object pooling service, and

increases the performance of enterprise applications.

As shown in Figure 2, PATS is composed of three

major components: (a) monitoring interface which

passively receives incoming calls to SFCs; (b) clients’

behavior pattern modeling engine which builds each

client’s behavior pattern according to its callings on

SFCs; and (c) tuning analyzer which decides about

the size of each OPR according to built behavior pat-

terns.

Figure 2: The internal architecture of PATS. Monitoring

interface component provides the clients’ behavior pattern

modeling engine with incoming requests from each client.

Then this engine according to the sequences of incoming

requests, builds and updates the behavior pattern of each

client as a separate markov model. The tuning analyzer

component uses these markov models to generate efﬁcient

tuning signals which will be sent to the associated object

pooling service.

As our tuning solution, PATS monitors requests

sent from each client to SFCs. In a real-time man-

ner, it builds and updates the dynamic behavioral pat-

tern of each client according to requests being sent

to SFCs. Using those behavioral patterns, we pre-

dict the future workloads which are more probable

to be sent to different SFCs. Through the prediction

of workloads on each SFC, our proposed tuning ana-

lyzer dynamically tunes the size of available OPRs in

the object pooling service. However, the object pool-

ing service is responsible to apply new size restric-

tions deﬁned by the tuning analyzer (i.e. releasing

extra objects loaded to each OPR). Also, the tuning

analyzer reports the new size of each OPR in terms

of a fraction of the total size of memory available for

object pooling purposes. In other words, the size of

objects loaded to each OPR does not affect the tuning

computation process. In the next section, we present

our prediction approach which is based on DMM.

3 BEHAVIOR PREDICTION

We are interested in building models of behavior

which are able to support the prediction of future

actions. We deﬁne the sequence of callings each

client sends to different SFCs as its behavior. Markov

model as a simple, yet efﬁcient prediction mechanism

is widely used to capture the sequential dependence

and constraints of behavioral patterns (Galata et al.,

1999). According to the deﬁnition (Eirinaki et al.,

2005), the n-order Markov model is a directed graph

with attached probabilities to each of its edges. The

transition probabilities of each state in this graph only

depend on ‘n’ previous states.

Using a ﬁxed order Markov model needs us to

know the length of required memory for prediction

(Galata et al., 1999). However, because of the dy-

namic and variable nature of the workload that we are

going to model (i.e., the requests being sent to SFCs

by each client), using ﬁxed order Markov models is

not applicable. In contrast, a Variable Length Markov

Model (VLMM) (Stefanov et al., 2005) not only cap-

tures high-order temporal dependencies but also lo-

cally optimizes the required length of memory (Galata

et al., 1999).

Dynamic Markov Compression (DMC) (Cormack

and Horspool, 1987) has taken advantage of VLMM

to ﬁgure out the probabilities of occurrences of fu-

ture symbols within compressing/decompressing data

streams. Here we replace the symbols of data stream

with the calling sequence that each client sends to dif-

ferent SFCs, and using DMC approach we build a

Dynamic Markov Model (DMM) to predict the future

SFC callings (cf. Figure 3). So, our prediction engine

updates Markov models as follows:

1. Receives the next SFC request from monitoring

interface component.

2. Finds the target Markov model according to the

identity of client.

3. Updates the target Markov model using DMC

approach.

And for prediction purposes, we extract the re-

quired client’s associated model from the prediction

engine. The probability of each transition from the

current state shows the probability of calling on the

attached SFC to that transition. Therefore, the num-

ber of transitions from each state of our Markov mod-

els equals to the total number of SFCs associated with

the object pool service.

ICSOFT 2007 - International Conference on Software and Data Technologies

Figure 3: A sample DMM built according to calling se-

quences sent to SFCs. i.e. when the prediction engine is

in state 1, then it will be aware of the fact that the prob-

abilities of incoming calls for SFC

, SFC

, and SFC

are

10%, 70%, and 20% respectively. Consequently, when the

prediction engine sees the next call is SFC

, it moves on the

edge that is labeled SFC

. Then, the probabilities for next

incoming calls will be extracted from the new current state

of the engine.

4 AUTOMATIC TUNING

In PATS architecture, tuning analyzer component is

responsible for determining the optimal size of each

OPR for a given state of the system. This tuner

as an effective automatic tuning solution (Sullivan,

2003) meets the following design criteria with rea-

sonable time costs: running without human interven-

tion, adapting to workload variations, and responding

to previously unseen workloads.

To tune the size of each OPR, we must ﬁgure out

the probable workloads which are going to be gener-

ated by each client for a SFC connected to the given

OPR. For this purpose, we use the associated Markov

model with each client. Each Markov model adapts

itself with workload changes, so our tuner works con-

sidering workload variations. Also we can extract

behavior patterns from each Markov model, so our

tuner is applicable for unseen workloads. On the other

hand, because all the parameters of each Markov

model is being tuned automatically, no human inter-

action is required. Subsequently, we explain the de-

tails of our tuning approach to show its accuracy and

time costs.

To predict the most probable workloads which are

going to be generated by each client, we can use two

different approaches with different time costs:

1. Finding the most probable path (i.e., a SFC call-

ing sequence) with a speciﬁc length T, for a given

Markov model from the current state. This re-

quires to solve the following problem: given a

Markov model λ composed of states, each of them

having N transitions (i.e. the total number of

available SFCs), how do we choose a calling se-

quence S = S

, S

, ..., S

so that P(S|λ), the prob-

ability of occurence of the calling sequence with

length T, is maximized. As this search space

contains N

different solutions, searching all of

the space results in a time complexity of O(N

However, we can use greedy algorithms (Cormen

et al., 2001) which ﬁnds an optimum solution, not

the best, with time complexity of O(N ∗ T). This

latter in each state chooses the transition with the

highest probability, moves to the next state using

that transition and repeats this process T times.

Although the greedy algorithm is more efﬁcient

than the non-greedy one, it suffers from a major

issue: when it arrives at a state whose transition

with the highest probability is connected to the

same state, it does not move to other states any-

more and ignores the random nature of those tran-

sitions probabilities.

2. Generating a random workload, with a speciﬁc

length T, according to a given Markov model λ

composed of states each of them having N tran-

sitions (i.e., the total number of available SFCs).

This approach uses an algorithm to generate ran-

dom numbers (i.e., numbers which are mapped

to SFCs) with given probability distributions (i.e.,

probability distribution of SFC callings from each

state of the given Markov model.). There are

techniques available for this purpose such as the

alias method proposed by Walker (Westlund and

Meyer, 2002) with initialization time complex-

ity of O(N ∗ log(N)) and generation time com-

plexity of O(1). However, there is also a vari-

ation of Walker’s original model by Vose (Vose,

1991) whose initialization time complexity is lin-

ear O(N) and generation time is the same O(1).

Consequently, exploiting this latter to generate a

random workload with length T requires O(N +

T) time.

According to the time costs of the two approaches

mentioned above and the real-time nature of request

processing in application servers, we use the second

approach which offers a linear time complexity to pre-

dict the future workload of each client. Subsequently,

using the predicted workloads of all clients, we can

predict the Future Workload Ratio (FWR) of each

OPR as follows:

FWR(OPR

) =

∑

c=1

seqn

(OPR

)

C∗ T

where I is the index of an OPR, C the total num-

ber of clients, T the length of predicted workload

A PREDICTIVE AUTOMATIC TUNING SERVICE FOR OBJECT POOLING BASED ON DYNAMIC MARKOV

MODELING

sequences for each client, and seqn

(OPR

) the to-

tal number of requests within predicted workload se-

quence which is going to be sent to OPR

by client

Although now we can use FWRs as future infor-

mation to tune the size of OPRs, we need to take into

consideration past information that give us a feed-

back about past efﬁciency of the tuning service. Here

we propose a variation of IPCM (Sadaoui and Shari-

ﬁmehr, 2006) which tunes the size of object pools ac-

cording to the past information. IPCM calculates the

size of each OPR according to its past information

including Hit Ratio

(HR) and Activity Ratio

(AR)

through the following formulas:

BaseSize(OPR

) = AR(OPR

) ∗C

− HR(OPR

) ∗ AR(OPR

) ∗C

Size(OPR

) = BaseSize(OPR

)

∑

j=1

HR(OPR

) ∗ AR(OPR

) ∗C

where I is the index of an OPR, AR a function

which returns the activity ratio of a given OPR, HR a

function which returns the hit ratio of OPR, C the to-

tal size of the object pool, and N the number of OPRs

available in the object pool.

In fact IPCM relies on hit ratios to avoid the oc-

curence of starvation for each OPR, and uses activity

ratios to divide object pool capacity in a fair manner.

However, both of these two factors are related to the

past information gathered from object pool activities.

Therefore, we deﬁne a new concept namely Workload

Ratio (WR) which contains both Future Workload Ra-

tio (FWR) and activity ratio which from now on we

call it Past Workload Ratio (PWR):

WR(OPR

) =

λ∗ PWR(OPR

) + ω∗ FWR(OPR

)

λ+ ω

where λ and ω are parameters which separately

tune the effect of FWR and PWR on the total amount

of WR. Generally, we set equal values for both λ and

ω which means FWR and PWR have equal inﬂuence

on the calculation of WR. Subsequently, we replace

AR with WR in IPCM’s formulas as follows:

BaseSize(OPR

) = WR(OPR

) ∗C

− HR(OPR

) ∗WR(OPR

) ∗C

The percentage of all accesses that are satisﬁed by the

loaded objects into an OPR.

The percentage of all accesses that are processed

through a speciﬁc OPR.

Size(OPR

) = BaseSize(OPR

)

∑

j=1

HR(OPR

) ∗WR(OPR

) ∗C

Therefore, according to these new formulas, we

consider both past and future information to ﬁgure

out the size of each OPR. However, in order to use

these formulas efﬁciently we must address an impor-

tant issue. As a matter of fact, we need to determine

an appropriate interval between tuning of OPRs. For

this sake, at ﬁrst we evaluate the performance of our

workload prediction in terms of Root Mean Square

Error (RMSE):

RMSE =

∑

i=1

(WR

−WR

′

)

Where N is the number of OPRs available in the

object pool, WR

the predicted Workload Ratio for

SFC

, and WR

′

the calculated Workload Ratio accord-

ing to the generated workload during the last time

window of workload measurement. Then, we use a

tuning trigger based on a RMSE threshold that auto-

matically runs the tuning process on OPRs whenever

the calculated value of RMSE is bigger than a speci-

ﬁed threshold. In this way, we make sure that PATS is

working with a managed level of error.

5 EVALUATION RESULTS

We conduct our experimentations using synthetic

workload models deﬁned by RUBiS (Cecchet et al.,

2002) to evaluate both scalability and performance of

PATS. These workload models are designed accord-

ing to standards introduced in TPC-W (Garcia and

Garcia, 2003) which model an online bookstore (see

clauses 5.3.1.1. and 6.2.1.2 of the TPC-W v1.6 spec-

iﬁcation (Council, 2001)). We develop an applica-

tion with SUN’s JDK 1.5.0

08 containing 15 SFCs

based on RUBiS proposed benchmark architecture.

We use a machine equipped with PIV 2.80 GHz CPU,

1 GB of RAM. Also we use MSSQL Server 2000 as

the backend DBMS running on MS Windows Server

2003 Service Pack 1.

To evaluate the performance of PATS, we mea-

sure the average size of OPRs for PATS, IPCM, and

static approach (i.e. the size of all OPRs are equal

and ﬁxed.). We carry out this evaluation for different

numbers of clients generating different types of work-

loads for several times. Figure 4 shows the result of

this experiment for 10 clients (cf. Figure 4.(a)), 20

clients (cf. Figure 4.(b)), 50 clients (cf. Figure 4.(c)),

ICSOFT 2007 - International Conference on Software and Data Technologies

0 10 20 30 40 50 60 70 80 90 100

10%

15%

20%

25%

30%

Average size of

Object Pool Regions (OPRs)

(a)

0 10 20 30 40 50 60 70 80 90 100

10%

15%

20%

25%

30%

(b)

0 10 20 30 40 50 60 70 80 90 100

10%

15%

20%

25%

30%

Workload repetition

Average size of

Object Pool Regions (OPRs)

(c)

0 10 20 30 40 50 60 70 80 90 100

10%

15%

20%

25%

30%

Workload repetition

(d)

IPCM PATS Static

Figure 4: Comparing the average size of OPRs for PATS, IPCM, and static approach for different numbers of clients generating

different types of workloads for several times. Diagrams a, b, c and d show the result of this experiment for 10, 20, 50 and 100

clients respectively. In each test case, we keep the number of clients ﬁxed and only change the number of workload repitition.

As shown the efﬁciency of PATS increases after more repitition of workload models and takes over two other approaches

signiﬁcantly.

and 100 clients (cf. Figure 4.(d)). As shown the ef-

ﬁciency of PATS increases by passing the time. The

period of time in which the efﬁciency of PATS is in-

creasing is called the learning period. In other words,

PATS needs time to learn the workload model of each

client. So, after passing the learning period which re-

sults in DMMs with a stable prediction accuracy, the

performance of PATS takes over two other approaches

signiﬁcantly.

To evaluate the required learning period for differ-

ent workload models, we generate workloads based

on same workload models for repeated times and

measure the mean true prediction after each repeti-

tion. True prediction means that the prediction en-

gine predicted the next incoming call was going to be

sent to SFC

and the next incoming call did go to the

SFC

. The result of this experiment for twelve dif-

ferent workload models is shown in Figure 5. With

respect to acquired results, it can be seen that the ac-

curacy of prediction in most of test cases gets stabi-

lized after ten repetitions.

To measure the relationship between the maxi-

mum size of DMMs (i.e. the maximum number of

available nodes in DMMs) and the prediction accu-

racy of PATS, we generate workloads based on the

same workload models with different allowed maxi-

mum size of DMMs. Figure 6 shows the impact of in-

creasing the size of DMMs on the prediction accuracy

of PATS for twelve different workload models. Inter-

estingly, increasing the size of DMMs improves the

prediction accuracy only for a while. In other words,

the prediction accuracy of PATS becomes stable with

a limited maximum size of DMMs.

6 CONCLUSIONS AND

PROSPECTS

In this paper, we have presented a novel predictive

automatic tuning service (PATS) for object pool ser-

vices. Integration of PATS into enterprise applications

has been considered by offering detailed architectural

requirements. Our proposed approach is based on Dy-

namic Markov Modeling for the prediction of future

A PREDICTIVE AUTOMATIC TUNING SERVICE FOR OBJECT POOLING BASED ON DYNAMIC MARKOV

MODELING

0 20 40

10%

20%

30%

40%

50%

Workload #1

Prediction accuracy

0 20 40

10%

20%

30%

40%

50%

Workload #2

0 20 40

10%

20%

30%

40%

50%

Workload #3

0 20 40

10%

20%

30%

40%

50%

Workload #4

0 20 40

10%

20%

30%

40%

50%

Workload #5

Prediction accuracy

0 20 40

10%

20%

30%

40%

50%

Workload #6

0 20 40

10%

20%

30%

40%

50%

Workload #7

0 20 40

10%

20%

30%

40%

50%

Workload #8

0 20 40

10%

20%

30%

40%

50%

Workload #9

Prediction accuracy

Workload repetition

0 20 40

10%

20%

30%

40%

50%

Workload #10

Workload repetition

0 20 40

10%

20%

30%

40%

50%

Workload #11

Workload repetition

0 20 40

10%

20%

30%

40%

50%

Workload #12

Workload repetition

Figure 5: The prediction accuracy versus the number of workload repetition for twelve different workload models. In each

test case, a speciﬁc workload model has been repeated 50 times. During this repetition, the learner Markov model has been

evolving to prepare better prediction accuracy. The prediction accuracy shows the percentage of true predictions made by

each Markov model. As shown in this ﬁgure, the prediction accuracy of most of Markov models have been stabilized between

30% and 40% only after 10 repetitions.

0 500 1000

10%

20%

30%

40%

50%

Workload #1

Prediction accuracy

0 500 1000

10%

20%

30%

40%

50%

Workload #2

0 500 1000

10%

20%

30%

40%

50%

Workload #3

0 500 1000

10%

20%

30%

40%

50%

Workload #4

0 500 1000

10%

20%

30%

40%

50%

Workload #5

Prediction accuracy

0 500 1000

10%

20%

30%

40%

50%

Workload #6

0 500 1000

10%

20%

30%

40%

50%

Workload #7

0 500 1000

10%

20%

30%

40%

50%

Workload #8

0 500 1000

10%

20%

30%

40%

50%

Workload #9

Prediction accuracy

Maximum DMM size

0 500 1000

10%

20%

30%

40%

50%

Workload #10

Maximum DMM size

0 500 1000

10%

20%

30%

40%

50%

Workload #11

Maximum DMM size

0 500 1000

10%

20%

30%

40%

50%

Workload #12

Maximum DMM size

Figure 6: The prediction accuracy versus the maximum size of a DMM (i.e. the maximum number of available nodes in a

DMM) for twelve different workload models. As shown in ﬁgure above, increasing the size of DMMs improves the prediction

accuracy only for a while. In other words, using only a limited amount of memory space, Morkov models prepare a proper

level of prediction accuracy which is between 30% and 40%.

ICSOFT 2007 - International Conference on Software and Data Technologies

probable workloads. We have elaborated our tuning

approach and discussed its time costs. Finally, us-

ing RUBiS benchmark workloads, we have conducted

some experiments and evaluations. The results of our

evaluations prove the scalability and effectiveness of

using PATS in enterprise applications. The next goal

of our research is to investigate how context-oriented

approaches can be used to precise the workload mod-

eling. Another goal is to provide the capability to con-

sider the descriptive characteristics of workloads for

accelerating the process of workload model learning.

REFERENCES

Alur, D., Malks, D., and Crupi, J. (2001). Core J2EE Pat-

terns: Best Practices and Design Strategies. Prentice

Hall PTR, Upper Saddle River, NJ, USA.

Anderson, P. and Anderson, G. (2002). Enterprise Jav-

aBeans Components Architecture: Designing and

Coding Enterprise Applications. Prentice Hall Pro-

fessional Technical Reference.

Barrera, J. S. (1993). Self-tuning systems software. In Proc.

Fourth Workshop on Workstation Operating Systems,

pages 194–197.

Cecchet, E., Marguerite, J., and Zwaenepoel, W. (2002).

Performance and scalability of ejb applications. In

OOPSLA ’02: Proceedings of the 17th ACM SIG-

PLAN conference on Object-oriented programming,

systems, languages, and applications, pages 246–261,

New York, NY, USA. ACM Press.

Cormack, G. V. and Horspool, R. N. S. (1987). Data com-

pression using dynamic markov modelling. The Com-

puter Journal, 30(6):541–550.

Cormen, T. H., Leiserson, C. E., Rivest, R. L., and Stein,

C. (2001). Introduction to algorithms. MIT Press,

Cambridge, MA, USA.

Council, T. P. P. (2001). TPC Benchmark W, Standard Spec-

iﬁcation.

Crawford, W. and Kaplan, J. (2003). J2EE Design Patterns.

O’Reilly & Associates, Inc., Sebastopol, CA, USA.

Deshpande, M. and Karypis, G. (2004). Selective markov

models for predicting web page accesses. ACM Trans.

Inter. Tech., 4(2):163–184.

Ebner, E., Shao, W., and Tsai, W.-T. (2000). The ﬁve-

module framework for internet application develop-

ment. ACM Comput. Surv., 32(1es):40.

Eirinaki, M., Vazirgiannis, M., and Kapogiannis, D. (2005).

Web path recommendations based on page ranking

and markov models. In WIDM ’05: Proceedings of

the 7th annual ACM international workshop on Web

information and data management, pages 2–9, New

York, NY, USA. ACM Press.

Galata, A., Johnson, N., and Hogg, D. (1999). Learning

behaviour models of human activities. In Proc. British

Mashine Vision Conference (BMVC’99), pages 12–22.

Garcia, D. F. and Garcia, J. (2003). Tpc-w e-commerce

benchmark evaluation. Computer, 36(2):42–48.

Guo, F. and Solihin, Y. (2006). An analytical model for

cache replacement policy performance. In SIGMET-

RICS ’06/Performance ’06: Proceedings of the joint

international conference on Measurement and model-

ing of computer systems, pages 228–239, New York,

NY, USA. ACM Press.

Jordan, M., Czajkowski, G., Kouklinski, K., and Skinner,

G. (2004). Extending a j2ee server with dynamic

and ﬂexible resource management. In Middleware

’04: Proceedings of the 5th ACM/IFIP/USENIX inter-

national conference on Middleware, pages 439–458,

New York, NY, USA. Springer-Verlag New York, Inc.

Kircher, M. and Jain, P. (2004). Pattern-Oriented Soft-

ware Architecture: Patterns for Resource Manage-

ment. John Wiley & Sons.

Oberle, D., Eberhart, A., Staab, S., and Volz, R. (2004).

Developing and managing software components in

an ontology-based application server. In Middleware

’04: Proceedings of the 5th ACM/IFIP/USENIX inter-

national conference on Middleware, pages 459–477,

New York, NY, USA. Springer-Verlag New York, Inc.

Raman, R., Livny, M., and Solomon, M. (2003). Pol-

icy driven heterogeneous resource co-allocation with

gangmatching. hpdc, 00:80.

Sadaoui, S. and Shariﬁmehr, N. (2006). A novel object

pool service for distributed systems. In The 8th In-

ternational Symposium on Distributed Objects and

Applications, New York, NY, USA. Springer Verlag

New York, Inc.

Stefanov, N., Galata, A., and Hubbold, R. (2005). Real-time

hand tracking with variable-length markov models of

behaviour. In CVPR ’05: Proceedings of the 2005

IEEE Computer Society Conference on Computer Vi-

sion and Pattern Recognition (CVPR’05) - Workshops,

page 73, Washington, DC, USA. IEEE Computer So-

ciety.

Sullivan, D. G. (2003). Using probabilistic reasoning to

automate software tuning. PhD thesis, Harvard Uni-

versity Cambridge, Massachusetts. Adviser-Margo I.

Seltzer.

Sullivan, D. G., Seltzer, M. I., and Pfeffer, A. (2004). Using

probabilistic reasoning to automate software tuning.

In SIGMETRICS ’04/Performance ’04: Proceedings

of the joint international conference on Measurement

and modeling of computer systems, pages 404–405,

New York, NY, USA. ACM Press.

Vose, M. D. (1991). A linear algorithm for generating ran-

dom numbers with a given distribution. IEEE Trans.

Softw. Eng., 17(9):972–975.

Westlund, H. B. and Meyer, G. W. (2002). A brdf database

employing the beard-maxwell reﬂection model. In

Graphics Interface, pages 189–201.

A PREDICTIVE AUTOMATIC TUNING SERVICE FOR OBJECT POOLING BASED ON DYNAMIC MARKOV

MODELING