Mobile Agents

1
Mobile Agents
Laboratoire LSR Logiciels Systèmes Réseaux
Andrzej Duda LSR-IMAG Andrzej.Duda_at_imag.fr
2
Agents mobiles
Résumé

• Ce tutoriel présente la problématique des agents
  mobiles. Nous
• commencerons par une introduction des concepts et
  du problème de la
• migration de processus. Nous discuterons ensuite
  des problèmes
• de sécurité, de communication, de nommage, de
  localisation, de persistance.
• Ensuite, nous présenterons plusieurs systèmes
  existants Aglets,
• Odyssey, MOA, Voyager, et MAP. Pour illustrer des
  applications, nous
• donnerons quelques exemples, en particulier dans
  le domaine de
• l'informatique nomade et des réseaux actifs. Nous
  parlerons aussi des
• activités de normalisation au sein de FIPA et de
  OMG.

3
Overview

• Motivating example
• Principles of mobile code and mobile agents
• distributed execution and mobility
• What is an agent? Mobile Agent System?
• Issues in Mobile Agent Systems
• mobility
• security
• communication
• naming and locating

4
Overview

• Issues in Mobile Agent Systems
• agent management
• persistency
• resource consumption
• Examples of Mobile Agent Systems
• Applications
• nomadic access to WWW
• active networks - ANTS

5
Overview

• Emerging standards
• OMG
• FIPA
• Conclusion
• Resource pages
• Programming examples

6
Motivating example
Data
Network
Data
Data
7
Remote Database Access

• Connect a laptop via a wireless link
• Activate an agent
• Disconnect
• Agent
• query a directory for relevant databases
• move to database servers
• work on local data
• return to a given home location
• Reconnect and get results

8
Characteristics

• Operations on remote servers
• some of them may depend on the previous ones
• Move execution closer to data
• reduce traffic
• results are usually smaller than data
• reduce latency
• local access to data
• Asynchronous operation
• no active entity at client site
• suitable for nomadic and mobile hosts

9
Principles of mobile code and mobile agents

• Distributed execution paradigms
• Remote procedure call (RPC)
• Remote evaluation (REV)
• Code on demand (COD)
• Mobile agents (MA)

10
RPC

• Paradigm
• application is split into two parts
• server provides a fixed interface
• synchronous operation
• invocation, execution of the procedure, result

11
Example
REQUEST URL of an object
REPLY object MIME
Content-type image/gif
--- object content ---
Server WWW
Client WWW
httpd
Netscape
12
Characteristics of RPC

• Fine-grain remote operation with static and
  specialized interfaces
• SUN RPC
• object invocation - CORBA
• HTTP methods
• extension mechanism - CGI
• Suitable for LANs
• low latency
• sufficient bandwidth
• Synchronous - client waits for results

13
Remote evaluation

• Paradigm
• activity send a procedure to execute on the
  target site
• procedure groups several operations
• synchronous operation
• invocation, execution of the procedure, result

14
Example of REV - servlets
15
Characteristics of REV

• Coarse-grain remote operation
• client defines its own interface
• Suitable for WANs
• tolerates high latency
• reduces network traffic
• Synchronous - client waits for results

16
Code on demand
proxy
activity
Get

• Paradigm
• activity downloads a procedure
• local execution

17
Example of COD - applets
18
Characteristics of COD

• Coarse-grain remote operation
• may provide user interface
• Suitable for WANs
• tolerates high latency
• reduces network traffic
• Asynchronous - local execution

19
Mobile agents

• Paradigm
• activity executes on a site
• migrates to the target site
• resumes execution on the target site

20
Characteristics of mobile agents

• Coarse-grain remote operation
• Suitable for WANs
• tolerates high latency
• reduces network traffic
• Asynchronous
• activating site may disconnect
• Explicit notion of location
• proactive - an agent decides where to go
• reactive - responds to an event or request

21
Mobile code technology

• Execution mobility
• how to move code and execution across the nodes
  of a network?
• Data space management
• how to rearrange resources after migration?

22
Executing units
Computational environment
23
Executing units

• Executing unit
• sequential flow of computation
• Resources
• external entities (may be shared)
• files, objects, OS variables, other EU
• State
• execution state
• instruction pointer, stack/heap
• data state
• references to resources

24
Execution mobility

• Migrate the execution unit to another site
• Weak mobility
• code migration, no execution state
• some initialization data
• Strong mobility
• both code and execution state migration
• used for long time for other purposes
• load balancing (Condor)
• fault tolerance - checkpointing

25
Data space management

• Resource bindings
• by identifier
• binding to a specific resource
• by value
• resource of a specific type value does not
  change after migration
• by type
• binding to any resource of a given type

26
Mechanisms

• By move
• transfer resource

source CE
source CE
destination CE
27
Mechanisms

• Network reference
• distributed data space

source CE
source CE
destination CE
28
Mechanisms

• By copy
• create a copy of the resource

source CE
source CE
destination CE
29
Mechanisms

• By re-binding
• re-bind a resource of the same type

source CE
destination CE
source CE
destination CE
30
Taxonomy

• Common sense definition
• An agent is a program that a person or
  organization vests with its authority, that can
  run unattended for a long time (e.g., a week),
  and that can meet and interact with other agents.
  The person or organization is the agent's
  authority.
• One can distinguish two kinds of agent
• A stationary agent executes on one computer
  system. A mobile agent can execute on different
  computer systems at different times of its life.

Jim White, General Magic
31
Taxonomy of agents

• Program Agents
• Agents in Java
• Aglets
• Migrating Processes
• Telescript
• Shipped Scripts
• TCLAgents
• MAP
• Active Networks
• ANTS

• Intelligent Agents
• User Assistance Agents
• learning interface agents - P. Maes
• e-mail filters
• Multi-agent Systems
• e.g. KQML or ACL interactions
• collaborative systems

32
Types of agents

• Mobile Agents
• move between nodes
• Autonomous Agents
• own environment
• Isolated Agents
• do not collaborate with others

• Stationary Agents
• provide services
• Dependable Agents
• depend on their clients
• Collaborative Agents
• collaborate to accomplish a complex task

33
Mobile agents

• Delegation
• represent a user
• perform actions on his behalf
• Autonomy
• executes in its proper environment
• independent of the user
• Mobility
• extends autonomy to decide when to move, where to
  move, and how to move

34
Benefits of mobile agents

• Service customization
• easy upgrade to a new functionality
• Software deployment and maintenance
• automate installation process
• Overcome bandwidth limitations
• execution over unreliable and slow links
• Increased fault tolerance
• immune from partial failures checkpointing
• Suitable for nomadic computing
• asynchronous operation

35
First mobile agents

• Worms (Xerox PARC, 1982)
• distributed computations
• Enabled Mail(1994)
• active mail using SafeTcl
• Telescript (General Magic, 1994)
• electronic marketplace
• Messengers - M0 (Geneva, 1994)
• language à la Postscript

36
First Mobile Agents

• Tacoma (Tromsø, Cornell, 1995)
• system support for agents
• MAP (IMAG, 1995)
• agents in Scheme on WWW
• AgentTCL (Dartmouth, 1996)
• agents in TCL
• Agents in Java

37
Mobile Agents in Java

• Aglets (IBM, 1996)
• Mole (Stuttgart, 1996)
• JAE (Aachen, 1997)
• ARA (Kaiserslautern, 1997)
• MuBot (Crystaliz, 1997)
• Grasshopper (IKV, 1997)
• Concordia, (Mitsubishi, 1997)
• MOA (OpenGroup, 1997)
• Odyssey (General Magic, 1997)
• Voyager (ObjectSpace, 1997)

38
Mobile agents in Java

• New opportunities
• independence of hardware platform
• Java portable applications - "skinny client"
• Java run time, GUI, network support
• wide range of devices hand held, mobile
• availability of Java run time on many servers
• databases, groupware, WWW
• wrappers for legacy systems
• provide mobility to stationary applications

39
Why Java?

• Portability
• interpreted bytecode and Virtual Machine
• availability on many platforms
• Rich APIs
• AWT, JFC, JDBC, RMI, others
• component programming - Java Beans
• system, network, security support
• object serialization
• Development tools
• Movement of software industry to Java

40
Network support in Java

• Support for TCP, UDP et HTTP
• java.net.Socket
• java.net.DatagramSocket
• java.net.URL
• Connected mode java.io
• methods getInputStream() and getOutputStream()
• Specialize a socket via heritage class SSLSocket
  extends java.net.Socket

41
RMI - Remote Method Invocation

• Method invocation on objects from a remote Java
  Virtual Machine
• Distributed Java applications
• Interaction with interfaces (not with their
  implementation)
• (Un)marshalling object serialization
• References
• name serverURL rmi//java.sun.com/my_object
• argument or return value

42
RMI Architecture
Client
Server
Application
RMI System
43
Serialization

• Storage and transfer of objects
• Read/Write
• ObjectInputStream
• ObjectOutputStream
• transient and static not serilized
• Security
• qualifier private transient
• overloading read/write methods
• encryption

44
Why not Java?

• Resource management
• no support for managing memory and disk limits
• Process management
• no support for preserving the thread execution
  context
• weak mobility - agents in Java do not migrate,
  rather a thread invokes a specified method

45
Mobile agent systems
agent
security
place
agent
data
services
network

• Agent primitives
• move
• meet
• send/receive
• locate
• checkpoint

agent
security
46
Weak mobility

• Java no support for preserving the thread
  execution context
• Simulating process migration
• serialization of an agent
• pack all objects of an agent into a message
• save the state of opened channels (MOA)
• transport the message to the destination node
• check for security credentials
• deserialize
• reconstruct the object graph of the agent
• execute a specified method

47
Mobility of Aglets

• Dispatch method
• serialize the aglet
• call onDispatching before departure
• call onArrival after arrival

48
Structure of Aglets
onArrival()
onActivation()
run() ... if(!_activated) // aglet before
deactivation deactivate() if(!_remote)
// aglet before dispatching ... dispatch()
deactivation/ activation
dispatching
onDispatching()
onDeactivating()
49
Other systems

• Concordia
• specify itinerary
• sequence of sites and methods to execute
• before moving
• put partial results into a serializable object
• Odyssey
• when an Odyssey agent is transported from one
  system to another, the agent's thread is
  restarted at the destination
• worker
• set of tasks per destination

50
Strong mobility

• Specific language and operating system support
• save the state, move to a remote site, resume
  execution
• Telescript - proprietary language and OS
• MAP - modified Scheme interpreter
• Strong mobility in Java
• modification of the Virtual Machine (Sirac, ARA,
  Sumatra)
• save/restore the execution state of one thread
• preprocess agent programs (U. Darmstadt)
• limitations (multiple threads, libraries)

51
Strong mobility

• Problems
• multiple threads
• shared objects
• synchronization
• re-binding resources

52
Security

• Basic aspects
• secrecy
• protection of information
• authentication
• access control
• non-repudiation
• proof of an operation
• integrity
• non-modification of the information

53
Security

• Protecting hosts against malicious agents
• access control and isolated executing environment
• Protecting agents against malicious hosts
• hard problem - open cryptographic issue

54
Security

• Agents execute in a controlled environment
• sandbox
• Agents cannot carry encryption keys

data space
Executing unit
Computational environment
55
Java security support

• Current security support for applets
• typesafe language
• no pointers
• true arrays with bound checking
• bytecode verifier - looks for
• pointers,
• instructions or call methods with invalid
  parameters,
• non-initialized variables

56
Java security support

• Current security support for applets
• class loader
• system classes cannot be replaced
• system classes have their own security measures
• e.g. the file access class has a list of files
  that may be accessed by applets

57
New Java security support

• Java Security API in JDK 1.1
• KeyPairGenerator Class
• generate pairs of public and private keys
• MessageDigest Class
• calculate the message digest of specified data
• Signature Class
• sign and verify digital signatures
• Encryption and decryption in future Java
  Cryptography Extension

58
Host security
authority secret key
authentication
message digest
digital signature
encryption
authentication
agent system secret/public key
59
Host security

• Accept only authorized, not modified agents
• Authentication and integrity
• compute the message digest (MD5) on the content
  of the agent,
• encrypt the digest, the identity and the
  authority of the agent using a secret key
• concatenate with the agent
• receiving agent system decrypts the signature
• either using the shared secret key
• either using the public key of the sender
• compute the digest

60
Host security

• Secrecy
• encrypt the content of the agent
• either using the shared secret key
• either using the public key of the receiver
• Denial of service attacks
• limit resource consumption per agent

61
Agent security

• Protect against tampering by a malicious host
• change the lowest price found by the agent
• Conceal the agent program
• find the strategy used by the agent
• Digitally sign a document
• intercept encryption keys

62
Agent security

• Smartcards as a Safe Haven
• execute security sensitive code fragments on a
  Smartcard
• Executable encrypted functions
• encrypt a function
• execute it on an untrusted host
• decrypt the result
• (see C. Tschudin)

63
Digital signing by an agent

• Signing with a concealed function s(x) S(x) is a
  routine implementing s(x)
• execute y P(x)
• compute signature z S(y)
• output (y, z)
• Signing with a composition h s ? f
• execute y F(x)
• compute signature z H(y)
• output (y, z)

64
Security in some systems

• Aglets
• security manager and Java Security API
• Concordia
• security manager that uses security credentials
  of an agent
• for full deployment, it may use strong security
  functions from external authorities such as
  Verisign
• MOA
• compatible with JDK 1.1, not yet implemented
• JAR format with digital signatures

65
Communication

• Three types of communication
• mobile agent - stationary agent
• mobile agent - mobile agent
• rendez-vous of agents - locating and notifying
• communication language (KQML, ACL)
• mobile agent - client application
• messages
• wait until an agent returns

66
Communication

• Nature of communication
• locally
• come to a place to communicate - meet (Odyssey)
• remotely
• communication channels (MOA)
• channels migrate with an agent
• synchronous and asynchronous messages (Aglets)

67
Naming

• Unique identifier
• example in MOA
• hostportfamily_lounch-no.generation-no
• needed to control agents
• kill, suspend, resume
• or to communicate or synchronize

68
Locating

• Based on name servers
• lookup, register, unregister
• Different location schemes (MOA)
• update the home location of every move
• register at a predefined name server
• search based on a predefined itinerary
• forward based on the trails left during migration

69
Agent management

• Controlling agents at all stages
• creation, duplication, destruction
• Examples of different types of control
• go - proactive migration
• retract - reactive migration
• an agent is forced to move
• clone - duplicate an agent
• destruction - how to kill proliferating agents?

70
Persistence

• Execution state is saved for migration
• checkpointing
• store the state to permanent storage
• restarting agents
• restore the state and resume execution
• reliable execution in spite of failures

71
Resource consumption

• Limits enforced on resources
• agents
• lifetime, places, hops, open channels, clones
• money to buy resources (M0)
• places
• lifetime, open channels, agents
• Negotiation
• when agent arrives to a node, it negotiates how
  many resources it can use
• Resources not supported by JVM cannot be limited
• size of VM, size of files

72
Aglets

• Aglet
• create, clone, dispatch, retract, deactivate,
  activate, dispose
• Context
• aglet's workplace a means for maintaining and
  managing running aglets
• Itinerary
• aglet's travel plan
• Identifier
• globally unique and immutable
• Messages
• synchronous and asynchronous message passing

73
Odyssey

• Agent class
• creation subclass of Agent or Worker class
• Worker set of tasks and destinations
• Agent system
• a platform that can create, manage, execute,
  transfer, and terminate agents
• Place
• a context within an agent system in which an
  agent execute
• Support for agents
• Tickets, Petitions, ProcessNames

74
MOA

• Agent
• move, meet, openChannel, sendMsg
• Place
• a stationary object
• Agent Environment
• Agent Control (manages system resources)
• Name Server (tracks agent locations)
• Mover (transfer agent state)
• Messenger (messaging between components)
• different trust domains
• one AE per JVM

75
Voyager

• Built on top of CORBA
• Agents
• Voyager.startup(7000)
• public void travel()
• moveTo("tokyo9000", "atTokyo"
• Persistent objects
• saveNow method
• Scalable group communication
• synchronous, one-way, future, multicast
• federated directory service

76
MAP - Mobile Assistant Programming
move
activate
agent
agent
MAP node
MAP node
data
services
data
services
clone
report
collector
get-result
report
agent
MAP node
MAP node
activation node
data
services
data
services
77
MAP primitives

• Activation
• start an agent on a chosen node, return a
  capacity for a result collector
• Get result
• retrieve results from the collector if all
  agents finished, tell it

78
Assistant primitives

• Move
• checkpoint the execution state
• transfer to a remote node
• resume execution
• Clone
• checkpoint the execution state
• duplicate the state
• resume execution
• Report
• store a message in a collector

79
Implementation

• Scheme
• fully fledged programming language
• interpreted
• the state of the interpreter can be saved
• WWW
• MAP node - a WWW server
• communication on top of HTTP
• access to the local space of HTML documents
• Operational prototype
• application front-end in Java

80
Properties

• Strong mobility
• modified Scheme interpreter
• Strong system semantics in presence of
  communication link and node failures
• atomic actions
• successful or not (an agent moves to another node
  or stays, but not the both)
• Agents - persistent processes
• execution terminates eventually

81
Example of an assistant
(define (message msg) (assistant-report
(string-append (assistant- identity) "
" msg))) (define (name node) (string-append
"clone_at_" node))
82
Example of an assistant
(define (go-over node-list) (if (null?
node-list) (assistant-exit) (if (not
(assistant-move (car node-list))) (begin
(message "move failed") (go-over
(cdr node-list))) (if (not
(assistant-clone (assistant-node)
(name (assistant-node)))) (begin
(message "clone failed")
(go-over (cdr node-list))) (if
(equal? (assistant-identity) "main")
(go-over (cdr node-list)) (message
"work"))))))
83
Example of an assistant
(define (main . args) (begin (message
"Start \"go over\"") (go-over args)))
scenario (main "kita.imag.fr" "delos.imag.fr"
"fidji.imag.fr" )
84
Applications

• Distributed information retrieval
• gather information from a set of sources
• Active networks
• programmable routers
• Network management
• decentralize management architecture
• Advanced telecommunication services
• dynamic reconfiguration and customization
• Electronic commerce
• negotiate and perform business transactions

85
Application nomadic access to the WWW
Netscape
Alycta
query
HTML
A
R
nomadic host
86
Low bandwidth connectivity

• GSM
• low bandwidth connection (9.6 Kbit/s)
• 141 KB page with graphics
• downloaded in 117 s (if cached)
• 285 s on Renater

Internet
GSM
Server WWW
MAP node
87
Low bandwidth connectivity

• Reduce bandwidth bottleneck
• asynchronous download by MAP agent
• transfer over Internet not included in the user
  response time
• content distillation - adapt quality of graphics
• JPEG, quality 5 ? 25 KB in 21 s
• JPEG, quality 20 ? 62 KB in 52 s

88
Adaptive degradation

• 3086 KB 1928 KB 1333 KB
• 100 62 43

89
Active networks

• Traditional packet networks
• passive packets transit through routers
• their operation is defined by a protocol, e.g. IP
  
• Active networks
• packet - active capsule
• miniature programs executed at each router
• router perform customized computations on the
  messages flowing through them
• it can modify the packet contents

90
Two views

• Programmable switches
• distinction between in-band data transfer and
  out-of-band management channels
• inject custom processing routines into the
  required routers - COD
• router examines the header and appropriate
  program is dispatched to operate on the content
• Capsules
• every message is a program composed of
  instructions
• basic computations on the capsule content
• can invoke built-in primitives

91
Common programming model

• Program encoding
• support mobility, safety, efficiency (Safe-TCL,
  Java)
• Common primitives
• changing header, payload, length
• access to node address, time, link status
• control packet flow forward, copy, discard
• Node resources
• bandwidth, processing capacity, storage
• routing tables, node's MIB

92
Applications of active networks

• Merging and distribution of information
• multi-user, multi-site applications
• fusing information from a large number of sensors
• users with different encoding and presentation
  requirements
• active Web caches
• store and execute programs to generate
  dynamically computed pages

93
Applications of active networks

• Active network management
• collecting and collating data
• implement sophisticated approaches to network
  monitoring and event filtering

94
ANTS

• A toolkit for building and dynamically deploying
  network protocols
• Capsule
• generalized replacement for a packet
• reference to the forwarding routine
• Code group
• collection of related capsules transferred as a
  unit
• Protocol
• collection of related code groups treated as a
  single protection unit

95
Structure of an ANTS node
applications
node cache
spawned threads
x
object-x
y
object-y
node thread
channel thread
code cache
routing table
ANTS capsules
96
ANTS classes

• Class Key Methods
• Node address, get, put, routefornode,
  delivertoapp
• Channel send, receive, node
• Application send, receive (upcall), node
• Capsule evaluate, length, encode, decode

97
Example of ANTS programming

• Mobile Data - capsules used by source to send
  data to the mobile host
• Register - a capsule to register forwarding
  information while roaming

98
Register capsule

• // on entry Node n
• // home home agent
• // next node at which to register
• // forward address to be registered
• // go to foreign and then home agent
• if (n.address() ! next)
• n.routefornode(this, next) return
• // insert a forwarding address
• n.put(src, forward)
• // after doing foreign, do home
• if (n.address() ! home)
• forward next
• next home
• n.routefornode(this, next)

99
Mobile Data capsule

• // look up forwarding record
• f n.get(dst)
• // if found, update our route
• if (f ! null) next f.node
• // and continue on our way
• if (n.address() ! next)
• n.routefornode(this, next)
• if (n.address() dst)
• n.delivertoapp(this, destport)

100
Standards

• OMG
• MAF (Mobile Agent Facility)
• MASIF (Mobile Agent System Interoperability
  Facilities)
• GMD Fokus, Crystaliz, General Magic, Open Group
• interoperability between agent systems written
  in the same language
• agent management, agent transfer, agent and agent
  system names, agent system types, location syntax

101
Standards

• FIPA (Foundation for Intelligent Physical Agents)
• Agent management
• ACL (Agent Communication Language)
• sort of KQML
• Agent/Software integration
• Applications
• Personal travel assistance
• Personal assistant
• Network provisioning and management
• Audio/Video entertainment and broadcasting

102
Conclusion

• Mobile agents
• new attractive paradigm for distributed computing
• benefits
• service customization
• software deployment and maintenance
• overcome bandwidth limitations
• increased fault tolerance
• suitable for nomadic computing

103
Conclusion

• Problems
• generic Java platform
• security
• need for high-level libraries
• scripting language for agents?
• Active networks
• rapid deployment of new protocols
• application specific processing

104
Resource pages

• Papers on mobile agents
• J. Kiniry, D. Zimmerman. A Hands-on Look at Java
  Mobile Agents, IEEE Internet Computing,
  July-August 1997
• M. Baldi, G.P. Picco, F. Risso. Understanding
  Code Mobility, IEEE Trans. Soft. Eng., 24(5),1998
• Proc. 1st Workshop on Mobile Agents, LNCS 1219,
  Berlin, April 1997
• Proc. 2st Workshop on Mobile Agents, LNCS 1477,
  Stuttgart, September 1997
• A. Duda, S. Perret Une architecture d'agents
  mobiles pour les réseaux de stations nomades,
  Proc. CFIP 97, 1997 ftp//ftp.imag.fr/pub/DRAK
  KAR/duda/cfip.ps

105
URLs

• Aglets
• http//www.trl.ibm.co.jp/aglets
• MuBot
• http//www.crystaliz.com
• Grasshopper (IKV, 1997)
• http//www.ikv.de/products/grasshopper
• Concordia
• http//www.meitca.com/HSL/Projects/Concordia
• Odyssey
• http//www.genmagic.com
• ANTS
• http//www.sds.lcs.mit.edu/activeware

106
URLs

• MOA
• http//www.camb.opengroup.org/RI/java/moa
• D. Milojicic et al. Mobile Objects and Agents
  (MOA), Design Implementation and Lessons Learned
  Proc. USENIX COOTS'98, April 1998
• Voyager
• http//www.objectspace.com/Voyager
• FIPA
• http//drogo.cselt.it/fipa
• MAF/MASIF
• http//www.camb.opengroup.org/RI/MAF

107
URLs

• Mobile agent bibliography
• http//pent21.infosys.tuwien.ac.at/Research/Agents
  
• http//www.cs.helsinki.fi/hhelin/agents
• Security - C. Tchudin
• http//www.docs.uu.se/tschudin/pub
• Active network research in Europe
• http//www.docs.uu.se/tschudin/research/an-europe
  .html
• Serveur WWW de la recherche en réseaux actifs en
  France
• http//www.loria.fr/festor/RAF/RAF.html