Issues in OR Parallelism

• Alternative branches are independent:
  • Same results are obtained in parallel or sequentially.
  • All branches can be safely explored in parallel.
  • Same number of branches explored
    (only if ``all solutions''! See later).

• Issues in or-parallelism: mostly implementation issues.
  (but granularity, side-effects, cuts, and aggregation predicates complicate things)

• System organization:
  • System comprises a collection of workers (processes/processors).
  • Each worker is an LP/CLP engine with a full set of stacks.
  • A scheduler assigns unexplored branches to idle workers.

• Main implementation problem: alternative bindings - efficiently maintaining different environments per branch (e.g., p and p in example).

• Solutions proposed:
  • Sharing (e.g. Aurora System [Lusk et al NGC90], PEPSys/ECLPS [ECRC], etc.)
  • Recomputation (e.g. Delphi model) [Clocksin NGC88]
  • Copying (e.g. Muse system) [Ali et al IJPP90]