A Cost Driven Approach to Information Collection for Mobile Environments Qi Han Nalini Venkatasubramanian Department of Information and Computer Science.

A Cost Driven Approach to Information Collection for Mobile Environments

Qi Han Nalini Venkatasubramanian

Department of Information and Computer ScienceUniversity of California-Irvine

•Quality of Service enhanced resource management at all levels - storage management, networks, applications, middleware

QoS Aware Information Infrastructure

QoS Enabled WideArea Network

BattlefieldVisualization

BattlefieldVisualization

Data servers

CollaborativeMultimedia Application

Mobile hosts

Data servers

Motivation

Advanced level of tetherless mobile multimedia services requires The development of a wireless network that supports

integrated multimedia services Focus of prior work

The development of intelligent network management middleware services that provides agile interfaces to mobile multimedia services

Our objective: to provide support for mobility and QoS management at the middleware layer independent of the underlying specific network architecture

QoS-based Resource Provisioning

Issues Effective middleware infrastructure that must adapt to changing

network conditions Resource provisioning algorithms that utilize current system

resource availability information to ensure that applications meet their QoS requirements

Additional Challenges In mobile environments, system conditions are constantly

changing Maintaining accurate and current system information is

important to efficient execution of resource provisioning algorithms

The Information Collection Problem

Goal To provide information good enough for resource

provisioning tasks such as admission control, load balancing etc.

Need an information collection mechanism that is aware of multiple levels of imprecision in data is aware of quality requirements of applications makes optimum use of the system (network and server) resources

while tolerating imprecision of the information

Collected Parameters Network link status, Data server capacity (Remote disk

bandwidth, Processor capacity), Mobile host status

Directory Enabled Network Information Collection

Provide directory service as an information base for QoS-provisioning algorithms

feasible servers for requests, available network and server resources Uses distributed probes to monitor traffic and collect dynamic load

state information

Directory Enabled Information Collection Information Acquisition Directory Organization and Manipulation Approximation and Cost Scalability: Hierarchical directory organization + Caching

Former Information Collection Approaches for Non-mobile Environments

Instantaneous snapshot based techniques (SS) Monitoring module samples residue capacity of network link

periodically and updates directory with latest value Static range based intervals

Partition link capacity into static intervals and update directory with the interval number

Throttle (TR) the directory holds a range-based representation of the monitored

parameter, with upper and lower bounds that can vary dynamically Time Series (MA)

time series models are used to predict future trends in sample values with some defined level of confidence

Challenges in Information Collection Problem for Mobile Environments

Inherent tradeoff between information accuracy and system performance

Solutions for non-mobile environments are not appropriate for mobile environments Increased dynamicity Constant change of client access points to fixed

network

Our Approach

Dynamic range-based representation

Mobile host Aggregation driven collection

Source and consumer-initiated triggers and updates

2 phase information collection process

Address the tradeoff between accuracy of directory information and the update overhead costs for mobile environments

Information Source

Information Mediator

Information Consumer

Information Collection Framework

Server selectionMobility management

QoS management Mobile QoS management

Location management Information collection

mobile host

fixedhost

server router

Information Repository

…

Components of Information Collection Framework

Information source Managed entities: server, link, mobile or stationary host…

Information consumer Consumers data collected from sources

Information mediator Decision point of the information collection

Information repository Holds system state information about sources

AutoSeC (Automatic Service Composition) Framework

Aggregate Mobility Model

Xregion

Region i

Mobile host j at (xj(t),yj(t))

Ymax

Xmax

regionX

XX max

dim

dimdim

)(,mod

)()(

X

i

Y

tyXi

X

txjtA

region

j

region

ji

Aggregation of Region i at time t:

Yregion

Resource Utilization Factor

Resource utilization factor for network links:

Resource utilization factor for servers:

;,

.,),,(

1),,(

rl

avail

n

rl

availBWBW if

otherwise ifnrlUF

BWBWnrlUF

Otherwise n)r,UF(s,

requested than greater capacities NIC and DBBFCPU available if

NICNIC,

DBDB,

BFBF,

CPUCPUMaxnrsUF

n

rsavailr

savailr

savailr

savail

,

,,

,}1111

{),,(

Generalized Aggregation Based Information Collection (Gen-ABIC)

Use a range R:=[L,U] to represent the monitored parameter

Phase 1: Derives the aggregate mobility patterns Utilizes the aggregation status and current resource utilization

status to adjust the collection parameters such as sampling frequency SF and range size R

Phase 2: Utilizes feedback from the sources (source-initiated triggers and

updates) and consumers (consumer-initiated triggers and updates) for further customization of the collection process

Information mediator

State Diagram of Information Collection Process

Directory service

Regular probing

RangerelaxationChange

confirmed

Range tightening

Changeconfirmed

Range adjustment

Current rangeCurrent range

Noise filtering

Valu

e o

ut o

f range:

source

-initia

ted trig

ger

Thrashingavoidance

Accu

racy

not e

nough:

consu

mer-in

itiate

d trig

ger

Information source Information consumer

New range

New range:consumer-initiated update

New range:source-initiated update

Cost Factors in an Information Collection Process

Regular sampling overhead Crs

Regular directory update overhead Cru

Source/consumer-initiated trigger overhead Cst and Cct

Source/consumer-initiated directory update overhead Csu and Ccu

consumer

mediator

source

Cr

s

Cst

Cc

tCcu

Cru

Csu

Directoryservice

Optimal Range Size to Minimize the Cost

To minimize the overall cost, a good range size is needed to reduce the need for further updates

To avoid source-initiated triggers and updates, R should be big enough Pst=Kst/R2 , Psu=Ksu/R2

To avoid consumer-initiated triggers and updates, R should be small enough

Pct=Kct*R , Pcu=Kcu *R

To minimize Cost :

3/1)(2

cucuctct

susustst

KCKC

KCKCR

cucuctctsusustst PCPCPCPC

The CDIC Algorithm

CDIC Algorithm( ) /* invoked periodically */ /* Phase 1: aggregation driven coarse-grained adjustment of parameters /* Compute host aggregation level; Compute resource utilization level; switch ( resource utilization) { case high: set SF and R to be minimum; case low: set SF and R to be minimum; case medium: increase/decrease SF and R based on current aggregation level; }

/*Phase 2: fine-grained adjustment of range size */ Calculate Kst, Ksu, Kct, Kcu based on monitored cost factors appropriately; Set R to be optimal which minimizes the cost.

Issues of CDIC

The model parameters such as Pst, Psu, Pct, Pct need to be monitored

Monitoring complexity affects the system performance to a great extent

User QoS may be compromised Utilizing mobile host aggregation status to drive the information

collection process could sacrifice some individual requests’ QoS, but overall system performance is improved

Optimized Cost Driven Information Collection (Opt-CDIC)

Further reduce communication overhead without sacrificing the overall QoS Selective triggering

Turn off consumer-initiated triggering Lazy sampling

Reduce sampling frequency when The number of source-initiated triggers in a

given period is less than a pre-determined value The range is relaxed to exceed a certain value

Simulation Environments Request model

Request arrival as a Poisson distribution Request holding time is exponentially distributed

Traffic model Uniform pattern Non-uniform pattern

Mobility model Incremental individual mobility model High mobility and low mobility

Four ScenariosHigh mobility Low mobility

Uniform traffic HM-UT LM-UT

Non-uniform traffic HM-NUT LM-NUT

Simulation Objectives

Analyze the impact of information collection mechanisms on the overall resource provisioning performance

Information collection mechanisms SS, SR, TR, Gen-ABIC, CDIC, Opt-CDIC

Resource provisioning algorithm CPSS (Comined Path and Server Selection)

Performance Metrics Request completion ratio Overhead involved Overall efficiency

Simulation Results (Comparison of SS, SR, TR, Gen-ABIC under HM-NUT)

Completion ratio Gen-ABIC shows the highest completion ratio SS, SR and TR exhibit similar completion ratios

Overhead Increases with the increase of the number of requests SS introduces the highest overhead, while Gen-ABIC

has the least overhead Overall Efficiency

Gen-ABIC shows the highest overall efficiency

Simulation Results (Comparison of SS, SR, TR, Gen-ABIC under HM-NUT)

Simulation Results (Comparison of SS, SR, TR, Gen-ABIC under HM-NUT)

Simulation Results (Comparison of Gen-ABIC, CDIC and Opt-CDIC under HM-NUT)

For completion ratio Gen-ABIC is marginally higher than Opt-CDIC, but

much higher than CDIC Decreases with an increase of the number of requests in

the system For overhead

CDIC is the highest, and Opt-CDIC is the lowest For overall efficiency

Opt-CDIC is the highest

Simulation Results (Comparison of Gen-ABIC, CDIC and Opt-CDIC under HM-NUT)

Simulation Results (Comparison of Gen-ABIC, CDIC and Opt-CDIC under HM-NUT)

Simulation Results (Comparison of Gen-ABIC, CDIC and Opt-CDIC under LM-UT)

Simulation Results (Comparison of Gen-ABIC, CDIC and Opt-CDIC under LM-UT)

Conclusions

Coarse assignment of collection parameters (e.g. SF and R) is adequate to render satisfactory completion ratios under most traffic workloads and mobility patterns

Optimization of turning off consumer-initiated triggers and lazy sampling help reduce overhead to a great extent without lowering the completion ratio

Therefore, Opt-CDIC is a desirable strategy to collect network and server information in mobile environments

Future Work

Enhance AutoSeC for mobile environments by integrating Opt-CDIC with the other resource provisioning algorithms

Develop a scalable information collection architecture suitable for wide-area environments that incorporates distributed directory services