Consistent first solution speedups in OR-parallel execution of logic programs
- Illinois Univ., Urbana, IL (USA). Dept. of Computer Science
The authors consider the problem of searching for one possible solution for a logic program in parallel. Most of the proposed parallel search strategies l = for logic programs fail to provide consistent linear speedups over sequential execution. The speedups vary between sub-linear to super-linear and from run to run. They present search strategies using priorities for OR parallel execution of logic programs when both consumer instance and pipeline parallelism is exploited using the REDUCE/OR process model. These strategies yield consistent speedups to first solution, which increase linearly with addition of processors. This is achieved by prioritizing the consumer instances of a goal literal. In addition the total number of nodes expanded during parallel search is kept very close to that in a sequential depth-first search. The authors describe a technique called delayed release to accomplish this. This is done by assigning priorities to subgoals that need to be evaluated, and initially exploring one branch at each OR node in the search tree. The unexplored OR alternatives are released in a delay manner. The authors present performance data on large OR parallel benchmark programs. The data demonstrates consistent linear speedups to first solution and the effectiveness of the scheme in reducing redundant work with efficient use of memory.
- OSTI ID:
- 5693102
- Country of Publication:
- United States
- Language:
- English
Similar Records
Parallel implementation of logic programs
Parallel execution of logic programs