CSAIL Publications and Digital Archive header
bullet Research Abstracts Home bullet CSAIL Digital Archive bullet Research Activities bullet CSAIL Home bullet

link to publications.csail.mit.edu link to www.csail.mit.edu horizontal line

 

Research Abstracts - 2007
horizontal line

horizontal line

vertical line
vertical line

Agent Organization and Request Propagation in the Knowledge Plane

Ji Li

Problem Statement

In designing and building networks like the Internet, we continue to face the problems of scale and distribution. Traditionally, network analysis, diagnosis and management have been done manually by a small number of people. With the increased scale, penetration, and distribution of the Internet, those traditional manual approaches requires an increasing number of people to be involved, and the management task itself becomes increasingly complex. Therefore, a central problem is to make the network self-knowledgeable, self-diagnosing, and perhaps in the future self-managing. The knowledge plane (KP) was proposed as a new higher-level artifact to make the network more intelligent [1]. At an abstract level, the knowledge plane gathers observations, constraints and assertions, and applies rules to these to generate observations and responses. At the physical level, it runs on hosts and servers within the network on which knowledge is stored.

In the Internet, network applications often need to maintain efficient connectivity graphs for various purposes. Examples include overlay networks, content distribution networks, end system multicast, peer-to-peer networks, publish/subscribe systems, etc [2,3,4,5]. For instance, routing overlays build their own routing graphs to route around congested paths with comparable or even better performance [2]; end-system multicast constructs the application-layer multicast tree to transmit video efficiently [3]; nodes in peer-to-peer networks probe each other to find nearby neighbors to improve lookup performance [4].

Currently each of those applications builds and maintains its own connectivity graph by probing latency, available bandwidth, loss rate between hosts actively, which often incurs significant cost in the network as redundant operations are performed by them individually. We believe that the introduction of the knowledge plane as a common infrastructure can help those applications construct more efficient connectivity graphs with greatly reduced maintenance cost. In particular, I propose to design an application-independent mechanism at the network layer that can be used for network management and applications to build application-specific efficient connectivity graphs.

My Approach

I propose to build a network knowledge plane (NetKP) for the network layer and on top of it, sub-planes for network management (sub-KPs), each designed to provide on category for network mangement funcationality. That is, the KP consists of the generic NetKP and multiple specialized sub-KPs. Both the NetKP and sub-KPs are composed of agents.

The NetKP provides valuable knowledge about the Internet to network management and applications in a scalable and efficient way. The knowledge in the NetKP includes network topology, and performance information (latency, bandwidth, etc). I will address problems including agent organization and network knowledge collection and dissemination.

On top of the network knowledge plane, we construct multiple sub-planes for network management. Each sub-KP focuses on one network service. For example, agents interested in the DNS form a sub-KP about DNS that helps diagnose DNS failures. In this work sub-KPs are for the purpose of network management, but I expect that sub-KPs may be used for other purposes, for example, a sub-KP for music. The connectivity graphs of sub-KPs are constructed using the knowledge provided by the NetKP and knowledge in their specific areas.

Both network management and network applications can benefit from the NetKP and the sub-KPs. I will prototype several sub-KPs and network management capabilities. In the case of network applications, I will conduct case studies on intrusion detection.

System Design

In both the NetKP and sub-KPs, there are three tasks in common: (1) knowledge plane organization; (2) request propagation and knowledge dissemination; (3) knowledge management. The first task, knowledge plane organization, refers to how agents in the KP discover each other and how agents organize themselves together. The second task, request propagation and knowledge dissemination, deals with how to propagate requests so that requests can be resolved quickly and efficiently and how to disseminate knowledge so that agents interested in that knowledge can receive it in a timely way. The first two tasks are tightly related to each other, because agent organization largely determines how requests and knowledge can be propagated. The third task, knowledge management, addresses the question of how agents manage their local knowledge and learned knowledge. The first two tasks are the focus of this work.

I propose a region-based agent architecture to organize the knowledge plane. Both the NetKP and sub-KPs consist of agents in their layers respectively. Agents are responsible for collecting, disseminating knowledge and resolving requests. Agents also manage a distributed knowledge base. The knowledge base is distributed among agents since the network knowledge is distributed and managed by different parties in the Internet.

Agents are organized into regions to achieve scalability and efficiency. The concept of the region is proposed as a new network feature to facilitate network organization [6]. In the NetKP, regions are often constructed following autonomous systems (ASes) or corporate networks, and agents in a large autonomous system may be divided into several regions. In sub-KPs, agents with similar interests form regions based on comprehensive criteria.

Summary

The ultimate goal of the knowledge plane is to build a new generation of network that can drive its own deployment and configuration, that can diagnose its own problems, and make decisions to resolve them. There are many challenging issues to achieve this goal, such as knowledge representation and utilization, trust and security, economic incentives, etc. This work is a step towards the knowledge plane.

Agent organization and request propagation are two key issues in the knowledge plane. By designing and implementing the NetKP and sub-KPs, and conducting several case studies on network management and applications, I hope to address the two problems, improve our understanding of the knowledge plane and motivate future research in this area.

References:

[1] D. Clark, C. Partridge, J. C. Ramming, and J. Wroclawski. A knowledge plane for the internet. In Proceedings of ACM SIGCOMM 2003, August 2003.

[2] D. G. Andersen, H. Balakrishnan, M. F. Kaashoek, and R. Morris. Resilient overlay networks. In Symposium on Operating Systems Principles, 2001.

[3] Y. H. Chu, S. G. Rao, and H. Zhang. A case for end system multicast. In Measurement and Modeling of Computer Systems, 2000.

[4] Gnutella. Http://gnutella.wego.com.

[5] V. Ramasubramanian, R. Peterson, and E. G. Sirer. Corona: A high performance publish-subscribe system for the world wide web. In Networked System Design and Implementation, 2006.

[6] K. Sollins. Designing for scale and differentiation. In ACM SIGCOMM 2003 FDNA Workshop, 2003.

vertical line
vertical line
 
horizontal line

MIT logo Computer Science and Artificial Intelligence Laboratory (CSAIL)
The Stata Center, Building 32 - 32 Vassar Street - Cambridge, MA 02139 - USA
tel:+1-617-253-0073 - publications@csail.mit.edu