Combinatorial Optimization by Decomposition on Hybrid CPU--non-CPU Solver Architectures

TL;DR

This paper presents a hybrid CPU--non-CPU solver architecture using problem decomposition to efficiently solve large combinatorial optimization problems, demonstrating significant runtime improvements and enabling solutions on smaller non-CPU hardware.

Contribution

It introduces a novel problem decomposition method tailored for hybrid architectures, enabling large graph problems to be solved with small non-CPU devices and improving classical algorithms' performance.

Findings

Maximum clique problem solved on large graphs using small non-CPU hardware.

Decomposition approach improves classical algorithms' runtime by orders of magnitude.

Small, specialized hardware can be competitive with CPUs for specific problems.

Abstract

The advent of new special-purpose hardware such as FPGA or ASIC-based annealers and quantum processors has shown potential in solving certain families of complex combinatorial optimization problems more efficiently than conventional CPUs. We show that to address an industrial optimization problem, a hybrid architecture of CPUs and non-CPU devices is inevitable. In this paper, we propose problem decomposition as an effective method for designing a hybrid CPU--non-CPU optimization solver. We introduce the required algorithmic elements for making problem decomposition a viable approach in meeting the real-world constraints such as communication time and the potential higher cost of using non-CPU hardware. We then turn to the well-known maximum clique problem, and propose a new method of decomposition for this problem. Our method enables us to solve the maximum clique problem on very large graphs using non-CPU hardware that is considerably smaller than the size of the graph. As an example, we show that the maximum clique problem on the com-Amazon graph, with 334,863 vertices and 925,872 edges, can be solved with a single call to a device that can embed a fully connected graph of size at least 21 nodes, such as the D-Wave 2000Q. We also show that our proposed problem decomposition approach can improve the runtime of two of the best-known classical algorithms for large, sparse graphs, namely PMC and BBMCSP, by orders of magnitude. In the light of our study, we believe that new non-CPU hardware that is small in size could become competitive with CPUs if it could be either mass produced and highly parallelized, or able to provide high-quality solutions to specific, small-sized problems significantly faster than CPUs.