CPU Extensions for Accelerating Combinatorial Optimization Problems

Combinatorial optimization (CO) problems are widespread in computer science, but only a small part of them can be solved in reasonable time. Heuristic algorithms are used to accelerate solving these, mostly NP-hard, problems like allocation or partitioning. This is a trade-off between speed and optimality, as there is no guarantee to find the best solution. In order to speed up heuristic algorithms based on local search, CPU extension modules are proposed: (i) A linked-list based representation of the CO problem is directly supported by a dedicated, internal list memory. (ii) The CPU instruction set is extended by list manipulation instructions. This allows for fast movement within a neighborhood of solutions. This movement is a frequent operation for algorithms based on local search. (iii) Additional modules like a pseudo random number generator or lookup tables support nonlinear functions in hardware. (iv) A status module supplies statistical information on solution quality and status of the algorithm. (v) In order to support future parallelization, an acceptance prediction module is introduced. This module accelerates pipelined calculations. The extension modules are designed to be integrated into a processor core.