Managing a Cloud For Managing a Cloud For Multi Agent SystemMulti Agent System

By,By,Pruthvi Pydimarri,Pruthvi Pydimarri,Jaya Chandra Kumar BatchuJaya Chandra Kumar Batchu

AbstractAbstract

We present a novel execution environment for multiagent systems building We present a novel execution environment for multiagent systems building on concepts from cloud computing and peer-to-peer networks. The on concepts from cloud computing and peer-to-peer networks. The novel environment can provide the computing power of a cloud for multi-novel environment can provide the computing power of a cloud for multi-agent systems in intermittently connected networks. We present the agent systems in intermittently connected networks. We present the design and implementation of a prototype operating system for design and implementation of a prototype operating system for managing the environment. The operating system provides the user with managing the environment. The operating system provides the user with a consistent view of a single machine, a single file system, and a unified a consistent view of a single machine, a single file system, and a unified programming model while providing elasticity and availability.programming model while providing elasticity and availability.

Existing SystemExisting SystemHence, the availability of a high performance computing infrastructure is Hence, the availability of a high performance computing infrastructure is

essential for the success of such systems. Existing high performance essential for the success of such systems. Existing high performance computing infrastructures such as clusters or cloud data centers may computing infrastructures such as clusters or cloud data centers may not be accessible in remote areas with intermittent network not be accessible in remote areas with intermittent network connectivity such as in military battlefields, disaster scenarios, etc., connectivity such as in military battlefields, disaster scenarios, etc., where powerful expert systems are needed for analytics, resource where powerful expert systems are needed for analytics, resource scheduling, and other applications.scheduling, and other applications.

  

Disadvantage: Disadvantage:   

• To give an less performance and storage space. Network traffic To give an less performance and storage space. Network traffic consumption also very high due to non concentrating on application consumption also very high due to non concentrating on application status. status.

  • It is not possible to build a scalable, high-performance distributed It is not possible to build a scalable, high-performance distributed

data-storage service that facilitates data sharing at large scale.data-storage service that facilitates data sharing at large scale.

  

  

Proposed systemProposed systemAutonomyAutonomy: A computation running on APMAC consists of a collection of : A computation running on APMAC consists of a collection of

agents running on the nodes of a network which communicate with agents running on the nodes of a network which communicate with each other to perform a set of tasks. Nodes can join and leave the each other to perform a set of tasks. Nodes can join and leave the network at any time.network at any time.

ReactivityReactivity: The agents react in response to inputs from the environment : The agents react in response to inputs from the environment in a timely fashion. The reactions can be published into the in a timely fashion. The reactions can be published into the environment or actuated through actuators. Our results show low environment or actuated through actuators. Our results show low latency between a stimulus and its response. The local state of an latency between a stimulus and its response. The local state of an agent changes in response to changes in the environment.agent changes in response to changes in the environment.

DeliberativityDeliberativity: The function embodied in an agent can be described by a : The function embodied in an agent can be described by a set of production rules that allow reasoning about changes in the set of production rules that allow reasoning about changes in the environment. The rules can be provided by a domain expert or can be environment. The rules can be provided by a domain expert or can be learnt from data.learnt from data.

InteractionInteraction: APMAC enables asynchronous communication of each : APMAC enables asynchronous communication of each agent with its environment as well as with other agents.agent with its environment as well as with other agents.

ModulesModules

Agent ModelAgent Model

Distributed Key-Value StoreDistributed Key-Value Store

PPDOS Implementation: Multi-agent MicrokernelPPDOS Implementation: Multi-agent Microkernel

Aggregate the StorageAggregate the Storage

1. Agent Model:1. Agent Model:

An agent running on APMAC has a predefined structure shown below (in An agent running on APMAC has a predefined structure shown below (in Java). Java). 

An agent is deterministic, i.e., it embodies a function in the mathematical An agent is deterministic, i.e., it embodies a function in the mathematical sense, and reactive, i.e., its state changes only in response to sense, and reactive, i.e., its state changes only in response to “events” triggered by the environment. The local variables of an agent “events” triggered by the environment. The local variables of an agent are partitioned into monitored variables, i.e., variables through which it are partitioned into monitored variables, i.e., variables through which it receives stimulus from the environment (indicated by invoking the receives stimulus from the environment (indicated by invoking the SUBSCRIBE method in the constructor), internal variables that it uses SUBSCRIBE method in the constructor), internal variables that it uses to store intermediate values, and controlled variables that record its to store intermediate values, and controlled variables that record its response to the stimulus and are published to the environment response to the stimulus and are published to the environment (indicated by registering to the operating system through invocation of (indicated by registering to the operating system through invocation of the REGISTER method). The update methods consist of rules that the REGISTER method). The update methods consist of rules that allow the agent to compute responses to environmental stimuli by allow the agent to compute responses to environmental stimuli by reasoning about it and are described by guarded commands that are reasoning about it and are described by guarded commands that are triggered by events in the environment.triggered by events in the environment.

2.2. Distributed Key-Value StoreDistributed Key-Value Store

For a distributed key-value store, we use Scalaris. It is an open source For a distributed key-value store, we use Scalaris. It is an open source distributed key-value store that supports transactions. Although, by distributed key-value store that supports transactions. Although, by default, Scalaris provides in-memory storage, through Tokyo-default, Scalaris provides in-memory storage, through Tokyo-Cabinet it can be configured to use persistent local hard disks. It Cabinet it can be configured to use persistent local hard disks. It implements the Paxos algorithm to replicate providing ACID implements the Paxos algorithm to replicate providing ACID properties with strong consistency. It uses quorum algorithms to properties with strong consistency. It uses quorum algorithms to recover from random crashes and node restarts and can run recover from random crashes and node restarts and can run uninterrupted even under intermittent node failures as long as a uninterrupted even under intermittent node failures as long as a majority of the nodes in the network are available. majority of the nodes in the network are available.

3.3. PPDOS Implementation: Multi-agent MicrokernelPPDOS Implementation: Multi-agent Microkernel

We have implemented PPDOS in Java. The PPDOS microkernel We have implemented PPDOS in Java. The PPDOS microkernel consists of a set of “system” agents that provides operating systems consists of a set of “system” agents that provides operating systems services and manage the deployment and execution of user agents. services and manage the deployment and execution of user agents. The PPDOS microkernel and the deployed user agents The PPDOS microkernel and the deployed user agents communicate asynchronously through the Scalaris distributed key-communicate asynchronously through the Scalaris distributed key-value store. Microkernel modules are collections of system agents value store. Microkernel modules are collections of system agents that together provide a certain type of service. The microkernel that together provide a certain type of service. The microkernel modules based on their functionality can be categorized into modules based on their functionality can be categorized into basic basic servicesservices, , resource management servicesresource management services, and , and agent execution and agent execution and deployment servicesdeployment services. We provide the algorithms implemented by . We provide the algorithms implemented by the operating systems services below; for lack of space we omit the the operating systems services below; for lack of space we omit the correctness proofs. The basic services are the core services correctness proofs. The basic services are the core services provided by the microkernel that are invoked by the user agents to provided by the microkernel that are invoked by the user agents to achieve their computation and communication goals. The Group achieve their computation and communication goals. The Group management services are the services of the microkernel that management services are the services of the microkernel that maintain consistency in the event of nodes joining and leaving the maintain consistency in the event of nodes joining and leaving the network. A group is a dynamic collection of nodes. We have the network. A group is a dynamic collection of nodes. We have the Agent Execution Management Services that provide functions like Agent Execution Management Services that provide functions like loading, unloading and migrating agents from one node to another.loading, unloading and migrating agents from one node to another.

4.4. Aggregate the Storage Aggregate the Storage

In most cloud deployments, the disks locally attached to the In most cloud deployments, the disks locally attached to the compute nodes are not exploited to their full potential. compute nodes are not exploited to their full potential. Most of the time, such disks are used to hold local Most of the time, such disks are used to hold local copies of the images corresponding to the running VMs, copies of the images corresponding to the running VMs, as well as to provide temporary storage for them during as well as to provide temporary storage for them during their execution, which utilizes only a small fraction of the their execution, which utilizes only a small fraction of the total disk size.total disk size.

ConclusionConclusionWe have presented the architecture and implementation of APMAC, a We have presented the architecture and implementation of APMAC, a

novel execution environment for running multiagent systems on ad-novel execution environment for running multiagent systems on ad-hoc networks and an operating system PPDOS for managing it. hoc networks and an operating system PPDOS for managing it. PPDOS provides a single machine, strongly consistent view of the PPDOS provides a single machine, strongly consistent view of the infrastructure with a single file system.infrastructure with a single file system.