Visualization for Parallel Performance Evaluation and Optimization

The main challenge lies in visually representing the mental models users apply in attempting to understand performance information. It is challenging to relate performance information back to the abstractions that the user understands, particularly when a performance view that appeals to one person's mental model may have little in common with models held by others. The solution is that the visualization techniques and methods used to construct graphic displays of the data must be closely integrated with the models of parallel computation the data represent.

In the following section, we discuss a paralel performance visualization model and an underlying theory for applying visualization principles. We then discuss visualization principles and scenarios.

A Parallel Performance Visualization Model

• Performance analysis abstraction
• Performance view
• Performance visual abstraction
• Performance display
• Performance visualization abstraction

The key point in the model is that the performance visual design can and should incorporate knowledge of the performance analysis abstraction very early on, providing the basis for performance interpretation in the final visualization. A performance visualization abstraction must be instantiated in a performance visualizer tool that implements performance views, displays, and their mappings using environment-specific graphics technology, based on underlying graphics libraries, toolkits, and other resources.

Performance Visualization Evolution

Recently, performance visualization research has shifted its emphasis to the technology required to generate application-specific performance visualizations. The three primary objectives of this recent work are: - to exploit the perceptual capabilitites of sophisticated graphics in the design of performance displays - to provide support for high-level performance visual abstractions and their instantiation through visualization languages, graphics libraries, and data visualization environments - to involve the user directly in prototyping and customizing performance displays so that meaningful performance visualizations can be readily constructed and evaluated.

Visualization Concepts and Principles

• Context
  • Perspective
  • Semantic Context
  • Subview mapping
• Scaling
  • Multidimensional and multivariate representation
  • Macroscopic and microscopic views
  • Macro/micro composition and reading
  • Adaptive graphical display
  • Display manipulation
  • Composite view
• User Perception and Interaction
  • Perception and cognition
  • Observing patterns
  • User interaction
• Comparison
  • Multiple views
  • Small multiples
  • Cross-execution views
• Extraction of Information
  • Reduction and filtering
  • Clustering
  • Encoding and abstracting
  • Separating information

Visualization Scenarios

• Processor Utilization, Concurrency, and Load Balance
  Kiviat diagrams, for example, depict relationships among multivariate performance data, such as the utilization of various resources in computer systems. To depict processor utilization, each processor is represented in the diagram by a separate radial axis, and its percentage utilization determines a point along this axis, from 0 at the origin and 100 % at the perimeter. Connecting these points by straight lines forms a polygon whose shape and size give a quick visual indication of overall efficiency, individual processor utilization, and the load balance across processors.

  Utilization data can be integrated over the lifetime of the computation and broken down into busy, overhead, and idle values, giving a quick and effective visual impression of load balance and overhead, but no insight into concurrency.

  This deficiency is the remedied by the timeline display, where the same information is now given as a function of time, so that specific periods of busy and idle activity on specific processors can be identified and correlated across processors.

  The Kiviat diagram and timeline display above demonstrate a tension that someone exists in performance visualization between views that try to convey impressions of overall performance behavior and those that support the presentation of detailed information. The visual abstraction of the Kiviat diagram can be extended to show a history of processor utilization values in the form of a "Kiviat tube", formed from two-dimensional Kiviat "slices" layed out in time along the third dimension.

• Critical Paths in Parallel Computation
  The large amounts of data in parallel programs makes it extremely difficult to identify the source of performance problems. The critical path is the longest serial thread, or chain of dependences, running through the execution of a parallel program. The execution time of the program cannot be reduced without shortening the critical path, and hence it is a potential place to look for bottlenecks. For parallel programs based on message passing, an appropriate visual abstraction for depicting the critical path is a minor modification of a spacetime diagram, since data dependencies are satisfied by interprocessor communications.

• Access Patterns for Data Distributions
  Parallel programming languages, such as High Performance Fortran or parallel C++ (pC++), incorporate data distribution semantics. Interprocess communication in such languages is implicitly determined by the data distribution, and hence the selection of the distribution is the programmer's main control over parallel efficiency. Thus, an appropriate analysis abstraction is the proportion of local versus remote data accesses required to support a given choice of data distribution.