Alva L. Couch

Categories and context in scalable execution visualization

Abstract To understand the behavior of processors in a large-scale asynchronous execution, we must often consider the global execution context in which each processor is immersed. Global context is best described by scalable execution views that do not change in format, size, meaning, or clarity as processors are added to an execution. One way to produce a scalable view is to categorize processors by behavior and display category statistics. Categorical views are particularly useful when there is an inverse mapping from an arbitrary view region to the subset of processors whose behavior was described in the region. Then the user can define new categories graphically by specifying subregions of views, using graphical attributes such as color and texture to depict category membership. The execution visualization tool Seeplex implements this form of category management to provide scalable execution views.

On observed reproducibility in network configuration management

Abstract A rigorous language for discussing the issue of configuration management is currently lacking. To this end, we develop a simple state machine model of configuration management. Observed behaviors comprise the state of a host and configuration processes accomplish state transitions. Using this language, we show that for one host in isolation and for some configuration processes, reproducibility of observed effect for a configuration process is a statically verifiable property of the process. Using configuration processes verified in this manner, we can efficiently identify latent preconditions that affect behavior among a population of hosts. Constructing configuration management tools with statically verifiable observed behaviors thus reduces the lifecycle cost of configuration management.

On the Algebraic Structure of Convergence

Current self-healing systems are built from “convergent” actions that only make repairs when necessary. Using an algebraic model of system administration, we challenge the traditional notion of “convergence” and propose a stronger definition with improved algebraic properties. Under the new definition, the structure of traditional configuration management systems is a natural emergent property of the algebraic model. We discuss the impact of the new definition, as well as the changes required in current convergent tools in order to conform to the new definition.

Dynamics of Resource Closure Operators

We propose a framework for managing resources via convergent operators. Operators represent their need for a resource to a designated resource closure operator that manages the resource. We evaluate a specific design for a resource closure operator by simulation and demonstrate that the operator achieves a near-optimal balance between cost and value without using any model of the relationship between resources and behavior. Instead, the resource operator relies upon its control of the resource to perform experiments and react to their results. These experiments allow the operator to be highly adaptive to change and unexpected contingencies.

A Theory of Closure Operators

We explore how fixed-point operators can be designed to interact and be composed to form autonomic control mechanisms. We depart from the idea that an operator is idempotent only for the states that it assures, and define a more general concept in which acceptable states are a superset of assurable states. This modified definition permits operators to make arbitrary choices that are later changed by other operators, easing their composition and allowing them to maintain aspects of a configuration. The result is that operators can be used to implement closures, which can in turn be used to build self-managing systems.

Portable execution traces for parallel program debugging and performance visualization

There is much interest in defining a standard for event traces collected from parallel architectures. A standard would support free data and tool sharing among researchers working on varied architectures. But defining that standard has proved to be difficult. Any standard must allow user-defined events and avoid or hide event semantics as much as possible. The authors propose a standard based on a declaration language, which describes how the raw event trace is to be translated into a normal form. Analysis tools then share a common interface to a compiler and interpreter which use the declarations to fetch, transform, and augment trace data. This concept is evaluated through construction of a prototype declaration compiler and interpreter.<<ETX>>

Integrated debugging and performance monitoring for parallel programs

Based on experiences with various debugging tools, the authors believe that the ideal tool represents several forms of integration: integration of scope from simple debugging to advanced performance monitoring; integration of techniques such as graphical and textual displays and real-time and postmortem analysis; connections between event monitoring and execution control; coordination of instrumentation in the parallel operating system with that in the tool; and a considered user interface throughout all components of the tool. Each of these forms of integration is briefly discussed.<<ETX>>

On system rollback and totalized fields: An algebraic approach to system change

Abstract In system operations the term rollback is often used to imply that arbitrary changes can be reversed i.e. ‘rolled back’ from an erroneous state to a previously known acceptable state. We show that this assumption is flawed and discuss error-correction schemes based on absolute rather than relative change. Insight may be gained by relating change management to the theory of computation. To this end, we reformulate previously-defined ‘convergent change operators’ of Burgess into the language of groups and rings. We show that, in this form, the problem of rollback from a convergent operation becomes equivalent to that of ‘division by zero’ in computation. Hence, we discuss how recent work by Bergstra and Tucker on zero-totalized fields helps to clear up long-standing confusion about the options for ‘rollback’ in change management.

Dynamic Consistency Analysis for Convergent Operators

It has been shown that sets of convergent operators with a shared fixed point can simulate autonomic control mechanisms, but many questions remain about this management technique. We discuss how an autonomous agent can reason about whether its convergent operators share a fixed point with the operators of other agents. Using a concept of time based upon operator repetition, we show that a failure to achieve convergence within specific time limits can be used as a probabilistic indicator of inconsistencies in local policy. We describe a statistical inference technique that determines if an agents promise strategy is feasible. The strengths of this technique are that it is both scale-invariant and exterior to the operators whose consistency is being evaluated.

Coordination and information exchange among resource management agents

We aim to reduce the overhead of information gathering and updates in a scenario of semi-autonomous resource management agents in a multi-agent environment. We investigate the importance of coordinating timing of events compared with the level of information exchange among the agents; and how this affects deviations from best-case management as predicted by theoretical models. We demonstrate that asynchronous interaction between management agents requires only a small amount of information exchange between the agents to achieve near-optimal results. This suggests that there are lightweight alternatives to centralized data gathering and planning to achieve autonomic service management.