Intelligent Agents -Agent Mobility and Cloning

1
Intelligent Agents -Agent Mobility and Cloning

• Katia Sycara
• The Robotics Institute
• email katia_at_cs.cmu.edu
• www.cs.cmu.edu/softagents

2
Information server
Host
Agent1
Network
Agent2
Idle Host
Host
Agent1
Network
Agent2
3
Mobile Agents Overview

• In some types of applications, it is useful to
  have mobile agents
• software agents that can migrate between machines
  to better utilize resources
• agents that reside on a mobile platform, and need
  to adjust to changes in the machine location
• Mobile agents can improve task performance and
  efficiency of using bandwidth and CPU
• Mobile agents require special services to allow
  for their application

4
Mobile Agents

• Programs typically written in a script language
  that may be dispatched from a client computer to
  execute remotely on a server
• sending a program without state to a remote CPU
  to execute there program possibly communicates
  by RPC with other CPUs and then terminates (e.g.,
  Java programs)
• sending a program that carries its execution
  state to a remote CPU program executes and then
  moves to a third CPU or returns. This is called a
  mobile agent

5
Mobile Agents (cont)

• The mobile agent execution environment is called
  the Agent Meeting Place. It allows
• passing parameters between the application
  program and the modules of the agent
• binding of user interface libraries of the
  clients operating system
• binding to other operating system functions such
  as the memory manager, the file system, etc, in
  particular to the message transport service
  (sending and receiving agents via the
  communication infrastructure)

6
Mobility Process

• Suspend current process or create daughter
  process
• Collect new or suspended process including
  execution state into a message expressed in
  machine independent form
• Send message to
• a particular known a priori machine
• a post office that performs address resolution
• intermediate designation that routes the agent
  based on message content (semantic routing)

7
Mobility Process (cont)

• Upon receipt of the message, the destination
  message subsystem delivers it to the servers
  agent execution environment
• In server execution environment, message is
  reconstituted into the execution thread(s) and is
  dispatched
• Execution continues at the next instruction
  following the instruction that initiated mobility

8
Mobility Process (cont)

• During execution at the server, the agent passes
  information from the client application and
  receives information from it, or
• Interacts with other agents on the server
• The agent may
• terminate its execution
• suspend its execution at the server waiting for
  some event from a server application or some
  other agent
• repeat the migration process, returning to its
  origin or going somewhere else, depending on the
  situation

9
What to Move?

• When agents collect and process information on
  the web, they consume bandwidth
• Attempting to reduce bandwidth consumption, a
  mobile agent can move to the information site
  instead of bringing the information to it
• But this requires bandwidth for code transfer, so
  if the agent is heavy weight, or it needs a
  little information - migration is not beneficial
• In practice, only when information transfer is
  significantly greater than code transfer
  migration is beneficial

10
Infrastructure Requirements

• A mobile agent that arrives at a foreign machine
  needs
• to be accepted by that machine be provided with
  access to machine resources
• to be written in a transportable language, able
  to run on any platform
• not to violate privacy or harm host machine
• be able to maintain state, to run itself, load
  code dynamically, stop itself
• protect itself from malicious host

11
What is the Price?

• Locally, a mobile agent needs to reason about
  itself, its needs, the network, and decide how to
  proceed
• Altogether, agent mobility imposes overheads
• in computation for decision making
• in infrastructure for receiving agents
• in bandwidth for migration

12
And What is the Promise?

• In the right settings, agent migration can
  improve performance significantly
• In some cases, mobility is unavoidable
• E.g., mobile platforms (laptops, palms)
• when a machine is about to disconnect from the
  network it is beneficial for an agent to
• get off the machine with essential information,
  state and processing
• continue living on the network until the machine
  is re-connected, possibly somewhere else
• move back to the machine, update it with changes,
  state

13
Promise (cont)

• Reduction of overall communication traffic over
  low bandwidth, high latency, high cost access
  networks employed by mobile computers
• Ability of the agent to engage in high bandwidth
  communication to search large free text databases
  on the server
• Ability of lightweight mobile computers to
  interact with heavy applications without prior
  detailed knowledge of the servers abilities

14
Promise (cont)

• Ability of the agent to integrate knowledge from
  the client and server while performing
  interfacing at the server
• Ability of the user to personalize services by
  customizing agents that are resident at a server

15
Current State

• Only a hand full of mobile agent systems support
  strong mobility, ie the ability to capture and
  transfer the full execution state of the agent.
• Since strong mobility requires capture of thread
  execution state, and since this is not currently
  possible in current Java VM, modifying the Java
  VM has been the strategy (e.g. NOMADS, ARA,
  Sumatra, WASP)

16
Current State (cont)

• In current Java Security Manager, there is no way
  to control
• -disc storage
• -number of I/O operations
• -number fo simultaneous print jobs
• -thread and process priorities
• -number of allowable open windows

17
Additional Mobile Agent Info

• Mobile agents list http//www.informatic.uni-stut
  tgart.de/ipvr/projecte/mole/mal/
• lists over 60 mobile agent systems
• voyager, oddyssey, aglets, Dagents are some
  mobile agent frameworks
• mobile agent languages safe-tcl, telescript,
  Java, Milog

18
Agent Cloning/Merging

• Cloning supersedes agent mobility
• Cloning could be done with or without the agents
  knowledge
• Cloning goals
• --To mitigate against agent overload
• --To achieve load balancing in a
  distributed system
• --To achieve higher throughput

19
Agent5 c5
Agent1 c1
Agent6 c6
T4
Agent7 c7
T9
T7
T5
T8
T
T2
T
T1
T
T3
Idle Host
Agent9 c9
T6
T10
Agent10 c10
20
Overloads in MAS

• Agents receive a stream of tasks T
• Tasks require capabilities and resources
• Agents capacity is limited. Therefore
• an agents resources for task execution may be
  insufficient. As a result
• agents may be overloaded
• at the same time others, on other machines are
  under-loaded
• the system utilizes resources inefficiently

21
Cloning

• Self-aware agents reason and anticipate overloads
  (predictive cloning)
• Self-aware agents detect overload after it has
  occurred (reactive cloning)
• Agents explore computational environment to find
  other agents, machines, resources
• Agents plan for task delegation, migration and
  cloning to resolve overloads
• Cloning is the creation and activation of other
  agents (on local or remote machines)

22
Create a clone
Agent5 c5
Agent1 c1
Agent6 c6
T8
T9
T7
Agent7 c7
T8
T8
Agent2 (clone)
T
T2
T1
T1
T
T3
Create a clone
T
T
Idle Host
Task delegation
Agent9 c9
T6
T9
Agent8 (clone)
Agent10 c10
23
The Cloning Algorithm(1) (performed by each
agent)

• Self awareness activity (do periodically, forever
  (in a separate thread of control))
• retrieve self available resources R(t0) (cpu,
  memory, disk space, comm. band width, etc)
• retrieve self current loads L(t0)
• anticipate current and expected Tl - capabilities
  and resources required, deadlines. Accordingly
• compute future (expected) loads L(t), resources
  R(t)
• compare R(t), L(t). Locate periods of overload
  OV(t) (what type of overload, what tasks cause
  it).

24
Algorithm(2) - Reasoning

• If OV(t) not empty, reason for cloning
• locate other agents and hosts
• for each located entity E
• retrieve current loads LE (t0)
• anticipate future tasks and loads LE (t)
• solve optimization problem to allocate tasks G in
  Tl that may cause OV(t) to entities. Refer to
• the remote loads LE (t)
• task deadlines and required capabilities,
  resources
• if E not agent - include creation/activation
  overhead

25
Algorithm(3) - Task Delegation

• The optimization process provides
• S, set of agents that can execute tasks in G
• for each ai c S, a unique, timed task list Tai
  (for each task in Tai, est and lst)
• H, set of hosts with available resources
• a timed task list TH for remote hosts
• For each ai c S,
• delegate tasks in Tai at appropriate time

26
Algorithm(4) - Create, Activate, Reason After

• While TH is not empty
• select a preferred host h in H
• create and activate a clone ah on h
• delegate tasks from TH to ah
• Reason, learn after cloning (subsequent slide)
• Reason about underloads. Consider
• merging with under-loaded agents
• self-extinction

27
Reasoning Guidelines

• Given the current task list and expected task
  distribution (provided or learned)
• If required resources at any time exceed agent
  maximum capacity-
• move tasks of the overloaded time to split-list
• if no agents can perform these tasks
• locate a host, create a clone, pass split-list

28
Reasoning for Cloning
29
Cloned-Agent Activity

• Introduces itself to MAS (via middle agent)
• Receives self task list Tl from creators TH
• Performs tasks (independently from creator)
• When Tl completed, if no new tasks arrived
  (reasons periodically w.r.t. underload)
• self-extincts, or
• merges with creator (or with another agent)

30
Reasoning After Cloning

• Collect info. about success, loads and
  availability of agents and hosts w.r.t. time
• Compute averages, variances and predictors of
  this information
• Use these results to update probabilities (of
  success of using agents/hosts) to the
  optimization problem
• Future cloning decisions will be affected

31
Cloning Simulation

• Setting
• 10 to 20 agents, up to 10 additional clones
• up to 1000 dynamically arriving tasks
• an agent can perform 20 tasks simultaneously
• Results
• up to 100 tasks - close to 100 of tasks are
  executed w/wo cloning.
• 100 to 350 tasks
• w/o cloning, MAS overall task execution poor
• with cloning approx. 85 of tasks met deadlines

32
System Throughput
44
33
Additional Utility of Cloning

• Can be used for agent migration and mobility by
• creating a clone
• passing all tasks to the clone
• original agent performs self extinction

34
Agent Merging

• When an agent or the underlying infrastructure
  detects underutilization of an agent
• kills the agent, if its capability is replicated,
  notifies its current customers, or
• selects agent with replicated capability and
  passes it the tasks of the underutilized agent,
  notifying its customers