A Survey of Computational Steering Systems and Future Directions

1
A Survey of Computational Steering Systems and
Future Directions

• Feng Xian
• fxian_at_cse.unl.edu
• Oct 29, 2003

2
Outline

• I. Definition of Computational Steering
• II. Problems in Computational Steering
• III. Survey of Computational Steering Systems
• IV. Future Directions

3
Definition
Computational steering can be defined as the
interactive control over a computational process
during execution (Marshall, 1990). Generally,
any interactive program can be viewed as steering
program. For example - Online debugging -
Interactive scientific simulation - Electrical
battlefield simulation Here, we mainly focus on
scientific simulation
4
Sequential way/Batch processing
Two simulation ways
Interactive way/Computational steering
5
Characteristics

• Close the loop of modeling, simulation and
  visualization
• Continuous visualization of intermediate results
• Interaction with simulation parameters
• Enables modifications of intermediate results
• User can directly influence the behavior of the
  simulation

6
Benefits

• Enhance productivity by greatly reducing the time
  between changes to parameters and the viewing of
  the results
• Enable the user to explore a what-if analysis
• The cause-effect relationships become more
  evident because changes in parameters become more
  instantaneous

7
Example Pollution Alert

• Problem an accident at a chemical plant resulted
  in the release of a toxic gas. Which regions need
  to be evacuated? Simulate the event faster than
  real time
• Some considerations on computational steering
• - A domain expert runs gas dispersion model
• - A meteorologist gives the whether forecast
• - They need to collaboratively steer the
  simulation and visualize the results in timely
  manner

8
General Approaches in implementing
Computational Steering Systems

• They operate in three phases
• Instrumentation Application code is modified to
  add monitoring functionality.
• Monitoring Program is run with some initial
  input data, the output of which is observed by
  retrieving important data about the programs
  state change.
• Steering Programs behavior is modified (by
  modifying the input) based on the knowledge
  gained during the previous phase by applying
  steering commands, which are injected on-line.

9
Outline

• I. Definition of Computational Steering
• II. Problems in Computational Steering
• III. Survey of Computational Steering Systems
• IV. Future Directions

10
Consistency

• Modifying/Steering parameters at unsafe
  pointers in the execution of the process might
  cause the computation to fail or produce an
  incorrect value

11
Scalability

• With the increase of steering components, the
  interaction overhead should not increase sharply.
  For example, the addition of distributed
  time-keeping technology to solve consistency
  problems typically results in great overhead to
  the program

12
Fault tolerance

• If a modification to a steering parameter failed,
  what action should the program take?

13
Latency

• Presentation lag
• The elapsed time between the occurrence of an
  event and the presentation of the associated
  graphic updates to user
• Steering lag
• The elapsed time between a steering action
  and the completion of the action

14
Outline

• I. Definition of Computational Steering
• II. Problems in Computational Steering
• III. Survey of Computational Steering Systems
• IV. Future Directions

15
Overview

• CUMULVS (Oak Ridge National Laboratory, 1995)
• CSE (Center for Math and CS, Netherlands, 1995)
• Progress (Georgia Institute of Technology,1995)
• Magellan (Georgia Institute of Technology,1997)
• SCIRun (University of Utah,1997)
• gViz (University of Leeds, NAG Ltd, HK, 2002)

16
Cumulvs

• CUMULVS is a middle-ware that allows parallel
  programs to easy incorporate real-time
  visualization and steering.
• Key features
• - a middleware/library built on PVM
• - user-directed checkpoints
• - provides a task immigration mechanism to
  achieve load balancing

17
Cumulvs Paradigm
Coordination of collection and dissemination of
information to/from parallel tasks to multiple
viewers.
Local Viewer Custom GUI
task
task
viewer
task
viewer
task
Remote Person Using VR Interface
Unix Host A
CUMULVS
task
task
task
task
task
viewer
task
NT Host B
Remote Person Using AVS
Unix Host C
18
Some mechanisms

• Coherence
• To prevent incoherent parameter updates,
  CUMULVS required users to acquire a token before
  modification.
• Fault tolerance
• Provides a fault-tolerant communication
  protocol so that failure in either an application
  or a viewer can be gracefully handled.
• When to make checkpoint
• At the iterative program. We commit a global
  checkpoint at the end of iteration.

19
Cumulvs Integration Interoperability

• Integration with InDEPS code development
  environment (ORNL SNL)
• Combustion Simulations
• Material Impact and Deformation
• Smooth Particle Hydrodynamics
• Tcl/Tk language bindings for Cumulvs (NCSA, ORNL)
• Viewers VTK Tango, VR / OpenGL viewer,
  Immersadesk, and CAVE
• Apps. Chesapeake Bay Simulation, Neutron Star
  Collision, DOD codes

20

• Pros
• Links to a variety of visualization tools
• Dynamic linking to running application
• Coordinated computational steering
• Fault tolerance

• Cons
• Non-transparent checkpoint, users need to
  identify where to place checkpoint
• Only applicable to iterative computations

21
CSE Architecture

• Key idea
• - a modular architecture for computational
  steering environment
• - provide a simple, flexible and minimal kernel

22
CSE Architecture
Researcher
expression
Rendering Satellite
Selection Satellite
Calculator Satellite
data
data
data
Data Manager
data
Simulation
23
CSE Architecture Description

• Data manager is implemented as a blackboard
• - manages a database of variables
• - acts as an event notification manager
• Satellite centers around the data manager
• - receives input data from users
• - subscribes to events that represent state
  changes in the data manager

24
Progress

• Steerable applications are developed through
  source-code modifications and steering is
  assisted by a run-time system.
• Key idea
• - Server-client model
• - provide three function hooks probe, sensor,
  actuator

25
Steering Object Operations

• Steering object is the component for steering
• Some fundamental operations
• - register register the steering objects with
  steering server
• - probe steering server read/write some
  non-critical variables in the application
• - sense forward the objects to the steering
  server for analysis
• - actuate performs a modification to the
  steering object.

26
Progress Architecture
Progress Client
Application
Sense Queue
Progress Runtime Server
Object1
Object1
Object1
Actuate Queue
Object Registry
Object2
Object2
Object2
Shared Memory
27
An Example Program
28
Magellan

• Derived from the Progress system Extends the
  client and server steering models.
• Uses a specification language ASCL to provide
  commands for monitoring and steering using the
  same objects as Progress.

29
SCIRun

• Key features
• - It is based on data flow programming model
  using boxes and wires approach.
• The main components are
• - module represents an algorithm or operation
• - port provides a connecting point to
  different
• stages of the computation
• - data type represents the conception behind
• the numbers
• - connection connects two modules together

30
(No Transcript)
31
gViz (Under development)

• gViz Aims to research and develop visualization
  middleware for e-science.
• Software components
• - IRIS (Commercial software, NAG Ltd.)
• - Globus Toolkit

32
Schematic of computational steering with gViz
33
Three main subdivisions of work

• Two grid-enable visualization systems, IRIS and
  pV3
• develop compression techniques suitable for
  transmitting very large streams of data
• build up appropriate XML languages for describing
  the structure and format of data to be analysed
  and for processing tasks.

34
(No Transcript)
35
Outline

• I. Definition of Computational Steering
• II. Problems in Computational Steering
• III. Survey of Computational Steering Systems
• IV. Future Directions

36
My tentative directions

• - Data management
• It common to generate hundreds or thousands of
  images, datafiles, and results. How to organize
  the data efficiently?
• Synchronization among collaborate users
• In the e-Science, how to make the views to
  different users coherent? Try to come up with an
  efficient scheme.
• - It can be used in p2p environment?
• Maybe or maybe not?

37
Thank you !!