A Program of Work for Understanding Emergent Behavior in Global Grid Systems Chris Dabrowski & Kevin Mills National Institute of Standards and Technology.

A Program of Work for Understanding Emergent A Program of Work for Understanding Emergent Behavior in Global Grid SystemsBehavior in Global Grid Systems

Chris Dabrowski & Kevin Mills National Institute of Standards and Technology

February 13, 2006

March/2006 2

What are emergent behaviors?

Why are emergent behaviors likely in global grids?

Can emergent behaviors be elicited or controlled?

How are NIST researchers investigating these questions?

Case study: denial-of-service (DoS) attack on simulated grid

Outline

March/2006 3

What are emergent behaviors?

Emergent behaviors are coherent system-wide properties

that cannot be deduced directly fromanalyzing behavior of individual components

Emergent behaviors typically arise in dynamic open complex adaptive systems,where system-wide behavior derives from

self-organizing interactions among myriad components

March/2006 4

Some Dynamic Open Complex Adaptive Systems

http://www.sover.net/~kenandeb/fire/hotshot.htmlhttp://www.sover.net/~kenandeb/fire/hotshot.html http://autoinfo.smartlink.net/quake/quake.htmhttp://autoinfo.smartlink.net/quake/quake.htm http://www.wtopnews.com/http://www.wtopnews.com/

http://www.avalanche.org/http://www.avalanche.org/

http://www.ics.uci.edu/relations/develop/rs2001/teitelbaum/sld012.htmhttp://www.ics.uci.edu/relations/develop/rs2001/teitelbaum/sld012.htm

http://www.english.uiuc.edu/maps/depression/photoessay.htmhttp://www.english.uiuc.edu/maps/depression/photoessay.htm

© M.F. Schatz and J.L.Rogers 1998© M.F. Schatz and J.L.Rogers 1998

© http://www.nationalgeographic.com/© http://www.nationalgeographic.com/ ©http://www.waag.org/realtime/©http://www.waag.org/realtime/

©http://emergent.brynmawr.edu 2003©http://emergent.brynmawr.edu 2003

March/2006 5

How might a complex system be detected?

fractal patterns

http://www.mbfractals.com/usergal/dougowen.html

fractal patterns

http://www.mbfractals.com/usergal/dougowen.html

self-similarity

http://www.physionet.org/tutorials/fmnc/node3.html

self-similarity

http://www.physionet.org/tutorials/fmnc/node3.html

ON = ParetoOFF = exponential

linear waveletsON = ParetoOFF = exponentialON = ParetoOFF = exponential

linear wavelets

1/f noise© J. Davidsen and H.G. Shuster 2000

1/f noise© J. Davidsen and H.G. Shuster 2000

http://complexity. orcon.net.nz/powerlaw .html http://heseweb.nrl.navy.mil/gamma/solarflare/24mar00.htm

power laws

http://complexity. orcon.net.nz/powerlaw .html http://heseweb.nrl.navy.mil/gamma/solarflare/24mar00.htm

power laws

Other ideas include: decrease in entropy or changes in statistical complexity

March/2006 6

What characteristics might lead to a complex system?

System Scale – order emerges from many interactions over space and time

Communications Locality – inability to know global state

Element Simplicity – inability to process all possible states

Feedback – elements can sense environment and estimate global state

Element Autonomy – each element can vary its behavior based on feedback

March/2006 7

Why are emergent behaviors likely in global grids? Scale: large number of clients and services interacting via indirect

coupling arising through use of shared resources

Communications Locality: clients cannot obtain complete and timely state of all resources – decisions must be made on partial information

Element Simplicity: clients possess limited processing power – decisions must be made with heuristics

Feedback: clients learn fate of resource requests and adapt subsequent requests based on updated information

Element Autonomy: clients decide how to proceed with no central control or direction

March/2006 8

Can emergent behaviors be elicited or controlled? Remains an open research question, for example:

– NASA exploring emergent programming to increase adaptability and survivability of future spacecraft (see Kenneth N. Lodding, “Hitchhikers Guide to Biomorphic Software”, ACM Queue vol. 2, no. 4) – MIT exploring amorphous computing where systems structure and specialize themselves from a common set of components (http://www.swiss.csail.mit.edu/projects/amorphous)

– Radhika Nagpal (Harvard) studying how to engineer and understand self- organizing systems (http://www.eecs.harvard.edu/~rad)

– Several researchers exploring application of economic mechanisms, such as markets, auctions, and present-value calculations, as means to elicit effective behavior in distributed systems

March/2006 9

How are NIST researchers investigating these questions?

1

51

101

151

201

251

301

351

401

451

501

551

601

651

701

751

801

851

901

951

1001

1051

1101

1151

1201

1251

1301

1351

1401

1451

S1

S7

S13

S19

S25

0

5

10

15

20

25

30

35

40

45

50

No

. Get

Tem

pla

te R

equ

ests

Time DSF

NSS 90 % Load

Goals • Understand self-organizing properties in service-oriented architectures (SOA)• Investigate mechanisms to shape emergent behavior in SOA• Improve related consortia specifications w.r.t. robustness, reliability, performance

Technical Approach • Apply modeling and analysis techniques from the physical sciences• Exploit exploratory data analysis and visualization methods• Investigate control techniques from biology and economics

Space-Time-State Evolution

Project Phases • Micro-model: 103 to 104 elements based on selected industry specs• Macro-model: 104 to 106 agent-based model containing selected abstractions validated against micro-model

March/2006 10

Architecture of Global Compute Grid

1. PUBLISH PROVIDER DESCRIPTION

2. DISCOVER PROVIDERS

PROVIDERSITE

CLUSTER

SERVER

SCHEDULERJOB

MANAGER

LOCALDIRECTORY

3. GET AVAILABILITY

4. NEGOTIATE USAGE

5. TRANSFER INPUT DATA

6. MONITOR EXECUTION

CLIENT

GLOBAL DIRECTORY

INTERNET

March/2006 11

Micro-model conception Layered Component Architecture

Network Layer: sites located in (x,y,z)-space used to compute distance in hops and simulate transmission delays;TCP-like simulated transport protocol; nodes model CPU delays, buffer & port capacity

Basic Web Services: WS- Addressing and Messaging WSRF: WS- Resource Property, Lifetime, Notification, Topics, Service Group Grid Services: MDS v4, WS-Agreement, and DRMAA

Major Grid Entities Service Providers: negotiate, schedule, execute, and monitor client tasks on

vector or cluster computers maintained at a related site Clients: discover providers, rank discoveries by earliest availability, seek

agreements, submit & monitor jobs

Client Grid Applications Application types: workflows of n sequential tasks, each with parallelizable sub-

computations – dependent tasks may not start until preceding task completes Tasks: types defined by tuple (required code, task parallelism, compute cycles)

and matched to processor component with suitable code and parallelism Workload: represented as a percentage of system capacity – regulated by

assignment of applications to clients

March/2006 12

Schematic showing operation of simulated grid

Scheduler

TaskControl

NegotiationControl

DSI

Grid Processor

Service Negotiator

Agreement

Grid ProcessorGrid Processor

DSIService

NegotiatorAgreement

Execution Control

CLIENT

Application

Client Negotiator

Task 1

DiscoveryControl

Task 2

Grid Processor

DSF

DSIService

Negotiator

Agreement

Client Negotiator

Task 3

TaskControl

NegotiationControl

ApplicationTask 1

DiscoveryControl

Task 2

Scheduler

DRMS Front-End

DRMS Front-End

DRMS Front-End

DRMS Front-End

DSF

spawnsspawns

negotiatesnegotiates

monitors

monitors

requests reservation

spawns

Supervisory Process Supervisory Process

spawns

GIIS

GRIS

GIIS

GRIS

GIIS

ProviderSite

ClientSite

Provider Site

March/2006 13

Case Study: DoS Attack on Simulated Grid• Deploy simulated topology: 200 nodes covering 30 provider sites and

12 clients, where each client uses one of two negotiation strategies• Negotiation strategies: serial reservation requests (SRR) or

concurrent reservation requests (CRR)• Run baseline: 50% workload for 200,000 simulated seconds and

measure the distribution of job completion times• Repeat run: inject service-provider spoofing with probability 50%,

effectively reduces system capacity by half on average• Repeat run: identical spoofing but introduce a strategy to resist

spoofing: identify spoofers and do not repeat interactions with them

1. Which negotiation strategy is more effective under normal conditions?2. Does the outcome change under attack?3. Does the outcome change when resisting attack?

Three Questions of Interest

March/2006 14

Bottom Line

1. CRR performs slightly better than SRR under normal conditions

2. CRR performs significantly better than SRR under attack scenario

3. Surprise: both CRR and SRR perform worse when resisting attackand the performance of CRR deteriorates more than SRR

The surprise arises because scheduling and execution of jobs inthe global grid is an emergent behavior arising from a self-organizing

property of distributed resource-management algorithms

March/2006 15

0.00

0.10

0.20

0.30

1000

0

2000

0

3000

0

4000

0

5000

0

6000

0

7000

0

8000

0

9000

0

1000

00

1100

00

1200

00

1300

00

1400

00

1500

00

1600

00

1700

00

1800

00

1900

00

2000

00

>200

000

Time

Pro

bab

ility

Comparative distribution of application completion times for two negotiation strategies (over 200+ repetitions)

Serial Reservation Requests (SRR) vs. Concurrent Reservation Requests (CRR) with No Spoofing

Concurrent Reservation Requests

Serial Reservation Requests

March/2006 16

Performance Degradation caused by Spoofing in Grid where 50% clients use SRR and 50% use CRR

0.00

0.10

0.20

0.30

Time

Pro

babi

lity

(a) No Spoofing

(b) Spoofing without Resistance

(c) Spoofing with Resistance (SURPRISE)

Comparative distribution of application completion times: (a) No Spoofing, (b) Spoofing without Resistance, and (c) Spoofing with Resistance (200+ repetitions)

March/2006 17

Decomposing performance degradation caused by spoofing

0.00

0.10

0.20

0.30

1000

0

2000

0

3000

0

4000

0

5000

0

6000

0

7000

0

8000

0

9000

0

1000

00

1100

00

1200

00

1300

00

1400

00

1500

00

1600

00

1700

00

1800

00

1900

00

2000

00

>200

000

Time

Pro

bab

ility

Spoofing without Resistance for CRR

Spoofing without Resistance for SRR

Spoofing with Resistance for CRR

Spoofing with Resistance for SRR

March/2006 18

0

50

100

150

200

250

3001 7 13 19 25 31 37 43 49 55 61 67 73 79 85 91 97 103

109

115

121

127

133

139

145

151

157

163

169

175

181

187

193

199

Time

Am

pli

tud

e

Two Time Series: (a) Reservations Created without Resistance and (b) Reservations Created with Resistance – 50% clients SRR and 50% CRR

Aggregate Reservations Created over Time under Spoofing with and without Resistance

(a) Without Resistance

(b) With Resistance

March/2006 19

Time Series for Application/Task Completions: Two Application Types without Resistance (lower blue) vs. with Resistance (upper red)

Time

Am

plitu

de

s

Time

Am

plitu

de

s

Serial Reservation Requests (SRR) Concurrent Reservation Requests (CRR)

Ap

plic

atio

n 1

Ap

plic

atio

n 2

Task2

Task1

Task3

Task2

Task1

Time

Am

plitu

de

s

Task2

Task1

Ap

plic

atio

n 1

Time

Am

plitu

de

sTask3

Task2

Task1Ap

plic

atio

n 2

Earlier

Earlier

Later

Later

Later

Later

March/2006 20

Conclusions Global Grids will be dynamic open complex adaptive

systems with self-organizing properties leading to emergent behaviors

Changes made to behavior in individual components could have pervasive and unexpected effects on global behavior

We need to develop a science of complex information systems in order to predict and control macroscopic behavior