Grand Challenge: A Distributed System That Automates The Planning
     Configuration, Control, And Operation Of Computer Networks

                          Douglas Comer
                Purdue University and Cisco Systems

Abstract

   As networks increase in size and complexity, the problems of network
planning, configuration, monitoring, troubleshooting, and security
enforcement have become critical.  Unfortunately, network management
remains the least understood aspect of networking.  Current tools focus
on the management of individual network devices and services or on specific
problems (e.g., detecting misconfigured routing or the analysis of flow
data) without treating a network  as a single, large system.  Consequently,
human managers must configure individual network entities in such a way that
the resulting network achieves desired behavior.  Because the consequences
of individual decisions are difficult to understand, minor changes in one
entity can have major, unintended impact on other parts of the network; the
complexity means that human managers cannot resolve all problems.  Indeed,
customers of network equipment vendors assert that the complexity of
managing their networks is limiting network size.  Automated network
management must be considered a crucial aspect of any future Internet.

New Abstractions

   Network management is analogous to operating systems in the 1960s:
vendors have built piecemeal solutions and found various ways to automate
individual tasks, but no consensus exists about underlying abstractions.
Furthermore, because future management will be automated, a management
infrastructure must supply abstractions that allow programmers to create
new management software instead of relying on the human interfaces found in
most existing commercial systems.  The challenge lies in finding a set of
abstractions for management analogous to the process, file, and device
abstractions invented for operating systems.

Why A Distributed System?

   The largest networks comprise multiple sites with many human managers
and multiple instances of management application software.  A management
system must provide coordination among sites to enforce global policy, while
allowing each site to have autonomy that permits operating during a period
of disconnection or making local changes that do not affect global policies;
such functionality requires a distributed system.

Research Problems Include

  * Find a small, elegant set of fundamental abstractions for network
    management that are sufficient for known management tasks, are
    extensible for new tasks, and accommodate management applications.
  * Devise a language that can be used to express management policies.
  * Create a distributed systems architecture for network management that
    can scale to arbitrary size; also consider issues of fault-tolerance,
    persistence, and temporary disconnection.
  * Formulate a role-based access control mechanism and associated roles
    sufficient to coordinate multiple managers and management applications.
  * Investigate the interface between a management system and underlying
    network elements to find a vendor-independent API that network elements
    can use to accommodate the new management architecture.
  * Define protocols for communication among components of the management
    system and network elements that are efficient and secure.

Resources Required

   Estimating 20 PIs and graduate students over 5 years plus distributed
lab facilities for experimentation, the cost will be $15M to $20M.