OLD PAGE FOR Scalable ABM Paper

Publications

Thesis Topics

  1. Effects of load imbalance
    • Both computation and communication
  2. Scalability of ABMs
    • Scaling potential for ABMs
    • Changing communication / computation ratio
    • Isoefficiency
    • Speedup potential
    • Could look at Gustafson's Law
    • etc
  3. “Cost” of conflicts and synchronization
    • Empirically analysis of the cost (in terms of execution time)
    • Could conduct statistical/empirical analysis on the average conflict per cell using average agent distribution

Supporting Topics

  1. Classify ABM's difficulty
    • easy to distinguish classifications (focus paper on “easier” categories - keeping in mind more difficult categories)
  2. Base implementation on Sugarscape model

Possible Experiments

  1. Test rectilinear decomposition strategy
    • Show the 'short cut' problem and demonstrate the mapping problem in recursive bisection
    • Show the load imbalance and length problem on dynamic row/col
    • Arrive that rectilinear decomp is the best
  2. Scaling experiment
    • Test the scalability of the implementation
    • Look at the changes in communication / computation
  3. Change sugar resources from regularly distributed to unevenly distributed (possibly clustered)
    • Test computational load imbalance
  4. Change sugar resources from regularly distributed to gridded (on the edge/corner of cells)
    • Attempt to maximize agent conflicts across cells
    • Test cost of conflicts
  5. Enable environment and agents to be “out of sync” for X number of iterations - then Sync
    • Test cost of syncronization
    • Must discuss tradeoffs of this
    • Another technique would be to compare sync/async MPI communication (but this is presumably less interesting to ABM people)

Conceptual idea

Literature

Analysis

Sugarscape

  • Agent movement (M)
    • Able to move NSEW (Von Neumann directions) only
    • Range limited by Vision (V) Values: 1 - (N-1) [ ghost buffer zone == V ]

Math for Sugarscape

Implementation

  • M (Movement)
    • Search for maximum sugar in NSEW directions with distance VISION FIXME (Currently move to random location)
    • Burn sugar according to distance moved * METABOLISM
    • Consume all sugar at new location
  • G (Growback)
    • Environmental (sugar) growback at GROWBACKA
  • R (Replacement)
    • If all sugar consumed in agent - die
    • Automatically replace agent with new agent in random location within local cell