Special Report: Network Provisioning

Network Science for High-Performance Networks There is a need for systematic scientific approaches to the design, analysis and implementation of the transport methods and to network provisioning. This calls for a new "science" of high- performance networking. In fact such a need has been recognized for future networks in a more general sense, wherein it is no longer sufficient to adopt ad hoc methods for their design and analysis. Several recent workshops [10-12] identified such fundamental and scientific aspects of future network research from a more general perspective (see Appendix B). While it is not the main goal of this workshop to identify the need for a science of networks in general, such need in the context of high performance networking has been strongly felt by both groups. With regard to DOE large-scale science applications, there are several essential components to such a science.

On-Demand Bandwidth and Circuit Optimization: Dynamic optimization and scheduling methods are needed to allocate the bandwidth pipes to various application requests. A comprehensive approach is needed for on-line estimation of the "bandwidths" of various network links and for their allocation on-demand to applications. It is likely that many of these allocation/scheduling problems are computationally hard, and efficient on- line methods must be designed to efficiently handle the allocations. Since the channel configurations are needed continuously across the network, the signaling aspects must be closely investigated to provide the required timeliness and reliability of the allocated channels. Once an understanding of signaling has been achieved, methods for both in- band and out-of-band signaling can be developed for dynamically switching the channels as per the allocations. There is a need for a scientific, systematic understanding of how to integrate the components for bandwidth allocation, channel scheduling, channel setup and teardown, and performance monitoring. At a higher level, a systematic framework is needed within which to analyze and classify application requirements and to design network configurations tailored to the resulting application classes.

Comprehensive Theory of Transport: A comprehensive theory of transport is needed to rigorously design transport methods tailored to the underlying provisioning modes. Such a theory should enable the rigorous evaluation and sound statistical analysis of the resulting transport methods. Such a theory would require a synergy and extensions of a number of traditional disciplines. Since delays and losses experienced by packets depend on the competing traffic, they exhibit apparent randomness. Such effects are particularly pronounced in heavily loaded networks. These effects are compounded by the complicated dynamics of TCP, which are apparently quite complicated even under very simple conditions. New stochastic control methods may be required to design suitable transport control methods. The resultant controllers are likely to be non-linear with delayed feedback, and new ideas in non-linear control theory may be required to analyze them. The feedback delays could themselves be a source of chaotic dynamics in non- linear systems. The area of statistics can play several roles in the theory by providing methods for designing rigorous measurements and tests. Optimization theory can potentially provide ways to obtain suitable parameters for tuning the various protocols.

Previous    Next    Table of Content for report: Network Provisioning    Home

Network Provisioning

 

 

 

Photuris.com - Optical Data Networking