Publish/Subscribe Systems on Node and Link Error Prone Mobile Environments Wireless and Mobile Systems Workshop

Publish/Subscribe Systems

on Node and Link Error-Prone

Mobile Environments

Sangyoon Oh, Sangmi Lee Pallickara,

Sunghoon Ko,

Jai-Hoon Kim, Geoffrey C. Fox

Community Grids Laboratory and Computer Science

Indiana University,

School of Information and Communication

Ajou University

Contents

z

Introduction

z

Cost model

z

Cost analysis

z

Performance comparison

z

Conclusion

Introduction

z

Advantages of

publish/subscribe systems in

mobile computing

•

Intermittent and high-latency

•

Decoupling publisher and

subscriber

•

Appropriate in mobile and

ubiquitous environments

•

Data dissemination services

•

Information sharing

•

Service discovery

ED

HHMS Proxy

ES

Wire d

W ired

ES (Event Source): Publisher ED (Event Displayer): Subscriber EBS (Event Brokering System): Server

Internet Cloud

ES

ES

EBS ES/_ED

ES/ ED

Radio tower

Cel lula

r

W ire

Motivations

z

Mobile environments are

error prone

•

Wireless link disconnection,

Power Exhaustion

z

In the presentation, we

analyze the influence of error

•

Mobile link and node

•

Comparison pub/sub to

client-server

and polling models

ES

EBS ES/

ED

Radio tower

ES

Cell_ular W

irele_ss

Node Failure

Wir eless

LA N

Node Failure Link

Failure

System model

Event Source

Event Brokering

System C_pub

Event Displayer

Event Displayer

Event Displayer

Csub

C

sub

C

sub

publish data (event)

System parameters

α

(publish rate)

β

(request rate or process

(reference access) rate)

c

_ps

(

α

)

(publish/subscribe cost per event)

c

_rr

(

β

)

(cost per request and reply)

c

_poll

(

α

,T)

(cost of periodic publish or polling)

s(n)

(effect of sharing among n subscribers)

t

_ps

(time delay for publish/subscribe)

t

_rr

(time delay for request and reply)

λ, λ

_s

, λ

_c

(failure rate of

communication link, server, and client)

μ, μ

_s

, μ

_c

(recovery rate of

Considerations and Assumptions

z

Analyzing four models:

Disconnection (failure) and reconnection (recovery) of link and node failure

1. Publish/subscribe

2. Event (broker) based request/reply

3. Conventional request/reply

4. Periodic polling

z

Cost metric

Æ

time delay to transfer message and additional time required due

to failure of communication links and nodes of servers and clients

z

Persistent files

for events are sharable among servers. When a server fails, the another

server can take over the role and publish/server model can continue its

transactions.

z

Durable database

Characteristics on Failures

Depends on system Depends on system

Wait until link reconnection any

periodic polling

Lost event or data during client failure Transaction needs

to restart after recovery Wait until link

reconnection, transaction needs to restart request

/reply request/reply

(RPC)

After recovery, client can access events

occurred during failure using events log

on durable database Server is

replicated and transaction can be

continued with sharable persistent

file Wait until link

reconnection, transaction is preserved request /reply Event message based request/reply

After recovery, client can access events

occurred during failure using events log

on durable database Server is

replicated and transaction can be

continued with sharable persistent

file Wait until link

Cost Analysis I

Publish/subscribe model

z

Cost of message transfer =

(c

_pub

+ n s(n)c

_sub

)

, plus delay time from

link and/or node failure

z

Failure of communication link

1. Probability of disconnection during transaction (μ/(λ+μ)(1- )

2. Probability of communication link disconnection (λ/(λ+μ))

t_pub + t_sub + {μ/( λ+μ) (1- ) + λ/(λ+μ)} {β/(μ+β)}(1/μ).

z

Failure of server

•

Assumption: 1) Another server takes over the failed server, 2) ignores the cost

required for transaction from failed server to backup server

t

_ps

= t

_pub

+ t

_sub

z

Failure of client

•

Assumption: Durable data Æ Client get any information after recovery (n_log)

•

Probability of i event occurred between failure and recovery Æ

•

If , i - n_logevents are lost due to exceeding limitation of capacity for event log

Thus, aver. number of lost events per client failure Æ

λ

ε−(t_pub+ t_sub)

λ

ε−(t_pub+ t_sub)

Aver. Delay time Probability of subscriber access

event before reconnection

Probability of disconnection

c c i

c α µ

µ µ α α + ⎟⎟ ⎠ ⎞ ⎜⎜ ⎝ ⎛ + c n c c c i n i c n i µ α µ α α µ α µ µ α α log log ) ( _log

1 ⎟⎟⎠

Cost Analysis II

Event (broker)-based request/reply

z

Cost of message transfer = c

_rr

, plus delay time from link and/or node failure

z

Failure of communication link

1. Probability of disconnection during transaction (μ/(λ+μ)(1- )

2. Probability of communication link disconnection (λ/(λ+μ))

t_rr+ {μ/(λ+μ)(1- )+λ/(λ+μ)}(1/μ)

z

Failure of server

•

Assumption: 1) Another server takes over the failed server, 2) ignores the cost

required for transaction from failed server to backup server

t

_rr

z

Failure of client

•

Assumption: Durable data Æ Client gets any information after recovery

•

Probability of i event occurred between failure and recovery

Æ

c c i

c α µ

µ µ

α α

+ ⎟⎟ ⎠ ⎞ ⎜⎜

⎝ ⎛

+

λ

ε

−trr

λ

ε

−trr

Cost Analysis III

RPC Based Request/Reply model

z

Assumption

•

No-persistent, No-durable

•

Additional overhead

Æ

‘useless computation’ (re-generating

request/reply, which is < t

_rr

+ t

_proc

)

•

‘Server recovery overhead’ is ignored like other models.

z

Failure of communication link

1. Link doesn’t fail

probability is 1 - μ/(λ+μ)(1- ) - λ/(λ+μ) Æt_rr

2. Link failed

probability is λ/(λ+μ) Ærecovery cost 1/μ + restart c_rr

3. Link failed during transaction

probability is μ/(λ+μ)(1- )

c_rr = [1 - { μ/(λ+μ) (1- ) + λ/(λ+μ)} ] t_rr

+ μ/(λ+μ) (1- ) { E(t_rr+t_proc,λ) + 1/μ + c_rr} + λ/(λ+μ){(1/μ)+ crr}

Î c_rr = [ [1 - {μ /(λ+μ) (1- ) + λ/(λ+μ)}] t_rr

+ μ/(λ+μ) (1- ) {E(t_rr+t_proc,λ) + 1/μ} + λ/(λ+μ)(1/μ)]

/ [ 1 - { μ/(λ+μ) (1- ) + λ/(λ+μ)]

λ

ε

−(tpub+ tsub)

λ

ε

−(tpub+ tsub)

λ

ε

−(tpub+ tsub)

λ

ε

−(t_pub+ t_sub)

λ

ε

−(t_pub+ t_sub)

λ

ε

−(t_pub+ t_sub)

λ

ε

−(t_pub+ t_sub)

Useless Computation: E(t_rr+t_proc,λ)

The function E denotes amount of

time, length x, and failure ratio of λ

E(t_rr+t_proc,λ) =

x x x t t x dt t _λ λ λ λ ε ε λ ε λε − − − − − − − = −

∫

₁ 1 ₁

Cost Analysis III

RPC Based Request/Reply model

z

Failure of server

1. Server doesn’t fail

: probability is1 - μ_s/(λ_s+μ_s)(1- ) - λ_s/(λ_s+μ_s) Æt_rr

2. Server failed

: probability isλ_s/(λ_s+μ_s) Æ recovery cost 1/μ_s+ restart c_rr

3. Server failed during transaction

: probability is μ_s/(λ_s+μ_s)(1- )

c_rr = [1 - { μ_s/(λ_s+μ_s) (1- ) + λ_s/(λ_s+μ_s)} ] t_rr

+ μ_s/(λ_s+μ_s) (1- ) { E(t_rr+t_proc,λ_s) + 1/μ_s+ c_rr} + λ_s/(λ_s+μ_s){(1/μ_s)+c_rr}

Î c_rr = [ [1 - {μ_s /(λ_s+μ_s) (1- ) + λ_s/(λ_s+μ_s)}] t_rr

+ μ_s/(λ_s+μ_s) (1- ) {E(t_rr+t_proc,λ_s) + 1/μ_s} + λ_s/(λ_s+μ_s)(1/μ_s)]

/ [ 1 - { μ_s/(λ_s+μ_s) (1- ) + λ_s/(λ_s+μ_s)]

z

Failure of client

•

Assumption

current data only and no-logging data

•

Only the data occurred during the failure will be lost.

s

λ

ε

−(tpub+ tsub)

s

λ

ε

−(tpub+ tsub)

s

λ

ε

−(tpub+ tsub)

s

λ

ε

−(t_pub+ t_sub)

s

λ

ε

−(tpub+ tsub) s

λ

ε

−(tpub+ tsub)

s

λ

ε

−(tpub+ tsub)

Useless Computation: E(t_rr+t_proc,λ_s)

The function E denotes amount of

time, length x, and failure ratio of λ

E(t_rr+t_proc,λ)=

x x x t t x dt t _λ λ λ λ ε ε λ ε λε − − − − − − − = −

∫

₁ 1 ₁

Cost Analysis IV

Periodic (Polling) model

z

Assumption

•

Polling is delayed until communication link is recovered.

•

the persistent file and durable database

z Link was disconnected or is disconnected during the transaction

•

Time delay =

t_poll(α,T) + {μ/(λ+μ) + λ/(λ+μ)}(1/μ)

•

Conceptual cost =

cpoll(α,T) + [1 – {μ/(λ+μ) + λ/(λ+μ)}] cdelay(α,T)

+ {μ/(λ+μ) + λ/(λ+μ)} c_delay(α,T+1/μ)

z

Failure of communication link

t_rr + {μ /(λ+μ)(1- )+λ/(λ+μ)}(1/μ)

z

Failure of server

t

_rr z

Failure of client

λ α

ε−t_poll( ,T)

c c i

c α µ

µ µ α α + ⎟⎟ ⎠ ⎞ ⎜⎜ ⎝ ⎛ + λ

ε

−trr

λ α

ε−t_poll( ,T)

λ α

ε−t_poll( ,T)

Performance Comparisons

by parametric analysis

z

Publish/subscribe system requires

less communication delay.

z

Assumption

System parameters

1, T, or α T

t_poll(α, T)

1 or 5 t_proc

1 t_ps , t_rr

0.05 - 0.1

μ_c

0.1

μ, μ_s

0.0001 – 0.5

λ , λ_s, λ_c

1/n - 1 s(n)

0, T, or α T

c_delay(α,T)

1 or α T

c_poll(α, T)

1 c_pub , c_sub

2 c_ps, c_rr

0.5

α, β

Values Param.

Only need

communication delay

ÅBackup server

Publish/subscribe system & Event-based

request/reply

Need additional time delay and lost

communication

ÅServer failure

RPC-based request/reply system

Performance comparisons II

0 0.5 1 1.5 2 2.5

0.0001 0.0005 0.001 0.005 0.01

Failure rate of communication links

Ti m e d e la y pub/sub pub/sub2 req/reply 0 0.5 1 1.5 2 2.5 3 3.5

0.0001 0.0005 0.001 0.005 0.01 Failure rate of server

Ti me d e lay pub/sub req/reply rpc

(

α

=0.5, t

_ps

=1, t

_rr

=1,

β

=0.5, and

μ

s=0.1)

(

α

=0.5, s(n)=1, t

_ps

=1, t

_rr

=1,

μ

=0.1,

and

β

=0.5 for pub/sub and req/reply

and

β

=0.2 for pub/sub2)

Performance comparisons III

0 0.5 1 1.5 2 2.5 3 3.5 4

0.0001 0.0005 0.001 0.005 0.01

Failure rate of link and server

Ti

me

d

e

la

y pub/sub_req/reply

rpc 0 2 4 6 8 10 12

0.05 0.1 0.2 0.5

Recovery rate of client

N u m b e r of l o s t e ve n ts n_log=0 n_log=10 n_log=20 n_log=30

(

α

=0.5, t

_ps

=1, t

_rr

=1, t

_poll

=1,

β

=0.5,

μ

=0.1,

μ

s=0.1)

Number of lost events per client

failure by varying recovery rate

of client

Experimental Setup

z

Using NaradaBrokering as message brokering system -- MOM

(Message Oriented Middleware)

for publish/subscribe

z

Using HHMS (Handheld Message Service) as primary

application level transport protocol

for publish/subscribe

between mobile device and conventional wired environment

z

_{Conventional RPC code using J2SE and J2ME MIDP 2.0}

_for

request/reply

z

_{Experimental Specifications}

•

Treo600:

PalmOS 5.2 144MHz ARM, 32MB, Sprint PCS Service (

<14.4kbps)

NaradaBrokering

z

Developed by Community Grids Laboratory of

Indiana University

z

Message Oriented Middleware (MOM)

•

Multiple protocol transport support:

In publish-subscribe Paradigm with

different Protocols on each link

•

Subscription Formats

•

Reliable delivery

•

Ordered delivery

•

Recovery

and

Replay

•

Security

•

Message Payload

options

•

Messaging Related Compliance

•

Grid

Feature Support

NaradaBrokering

Stream

NB supports messages

and streams

NB role for Grid is

Similar to

Handheld Messaging Service

z

Light-weight publish/subscribe message

service framework for mobile devices

z

Optimized application level transport

protocol using byte message format

Experiment results

z

Simple server

Æ

sending 10 bytes messages

z

Echo client

Æ

return the message back

z

Round Trip Time (RTT)

Å

median is chosen among 2000 iteration

z

_{It shows matching result with our parametric analysis}

t

_ps

= t

_rr

= 1

Delay time of sending message

1538.1 (t

_ps

)

89.7 (EBS – ES) 1448.4 (ED – EBS)

Pub/Sub

1330.6 (t

_rr

)

39.9 (gateway – server) 1290.7 (client – gateway)

RPC

Total (msec.) Wired

Conclusion

z

Advantages of publish/subscribe system in push

based mobile applications.

z

Problem: Error-prone mobile environments

•

Link disconnection

•

Power exhaustion

z

Publish/subscribe system has improved performance

and effectiveness on failure of client and server

node, and disconnection of communication link.

z

Our analysis shows publish/subscribe system is