Lecture 5 Sources of Parallelism and Locality in Simulation

Parallel Computer Model

Parallelism and data locality both critical to performance.

Basic Kinds of Simulation

Actually we can rank a system from discrete to continuous.

• Discrete Event Systems
• Particle Systems
• Lumped variables depending on continuous parameters
  • aka. ODE
• Continuous variables depending on continuous parameters
  • aka. PDE

Discrete Event Systems

1. Systems are represented as finite set of variables.
2. variable at a given time: state
3. each variable is updated by calling a transition function

System may be:

• synchronous: at each discrete timestamp evaluate all transition functions; also called a state machine
• asynchronous: transition functions are evaluated only if the inputs change
  • based on an “event” from another part of the system
  • also called event driven simulation

Approach:

1. Domain Decomposition - pick shape of domain:
   • Regular Meshes: minimize communication on mesh
   • Graph Partition: load balance + minimize communication
2. Run each component ahead using:
   • Synchronous: communicate each time stamp
   • Asynchronous: communicate only when necessary
     • Conservative: need deadlock detection
     • Speculative: need rollback if necessary

Others

Ignored in this page