A Novel CPU-GPU Cooperative Implementation of A Parallel Two-List Algorithm for the Subset-Sum Problem

Abstract

The subset-sum problem is a well-known non-deterministic polynomial-time complete (NP-complete) decision problem. This paper proposes a novel and efficient implementation of a parallel two-list algorithm for solving the problem on a graphics processing unit (GPU) using Compute Unified Device Architecture (CUDA). The algorithm is composed of a generation stage, a pruning stage, and a search stage. It is not easy to effectively implement the three stages of the algorithm on a GPU. Ways to achieve better performance, reasonable task distribution between CPU and GPU, effective GPU memory management, and CPU-GPU communication cost minimization are discussed. The generation stage of the algorithm adopts a typical recursive divide-and-conquer strategy. Because recursion cannot be well supported by current GPUs with compute capability less than 3.5, a new vector-based iterative implementation mechanism is designed to replace the explicit recursion. Furthermore, to optimize the performance of the GPU implementation, this paper improves the three stages of the algorithm. The experimental results show that the GPU implementation has much better performance than the CPU implementation and can achieve high speedup on different GPU cards. The experimental results also illustrate that the improved algorithm can bring significant performance benefits for the GPU implementation.well-known approach is the dynamic programming algorithm [7], which solves SSP in pseudopolynomial time, but it has exponential time complexity when the knapsack capacity is large. A tremendous improvement was made by Horowitz and Sahni [6], who developed a new technique that solves SSP in time O.n2 n=2 / with O.2 n=2 / memory space. The new technique is known as the two-list algorithm. On the basis of the two-list algorithm, Schroeppel and Shamir [8] proposed the two-list four-table algorithm, which needs the same time O.n2 n=2 / and less memory space O.2 n=4 / to solve SSP. Although many sequential algorithms have been designed to solve SSP in the past, Horowitz and Sahni's two-list algorithm continues to be the best known sequential algorithm.With the advent of parallel computing, a large effort has been made to reduce the computation time of SSP. Karnin [9] proposed a parallel algorithm that parallelizes the generation routine of the two-list four-table algorithm [8] using O.2 n=6 / processors and O.2 n=6 / memory cells in time O.n2 n=2 /. Ferreira [10] presented a brilliant parallel two-list algorithm that solves SSP in time O.n.2 n=2 / " / with O..2 n=2 / 1 " / processors and O.2 n=2 / memory space, where 0 6 " 6 1. Chang et al. [11] introduced a parallel algorithm that parallelizes the generation stage of Horowitz and Sahni's two-list algorithm [6]. They claimed that their parallel generation stage can be accomplished in time O..n=8/ 2 / with O.2 n=8 / processors and O.2 n=4 / memory space. On the basis of the generation technique of Chang et al., Lou and Chang [12] successfully parallelized the search stage of Horowitz and Sahni's two-list algorit...

Parallel implementation of the modified subset sum problem in CUDA

“…There are many algorithms developed to solve classic subset sum problem, some of them are: branch-and-bound, parallel two-list algorithm, genetic algorithm, etc. [4].…”

“…Their CUDA implementation will not show good speedup if the table does not fit within the device memory. Another approach [4] exploits the CPU-GPU cooperation in order to achieve a speedup factor of $9.2$ over the best sequential implementation. Our implementation is solely on GPU and does not use any table for dynamic programming.…”

Research and implementation of a high performance parallel computing digital down converter on graphics processing unit

“…The basic unit of execution in CUDA is the so called kernel. When a CUDA program invokes a kernel on the host side, these thread blocks within a grid are enumerated and distributed to multiprocessors with available execution capacity; all threads within a grid can be executed in parallel …”

Modern Platform for Parallel Algorithms Testing: Java on Intel Xeon Phi

“…GPGPU (General-Purpose computing on Graphics Processing Unit) is a typical instance. Example implementations: two-list algorithm for the subset-sum problem [9] or protein structure similarity search engine [10] illustrate the approach-parallel algorithms execution on GPU requires adjusting to the specific architecture.…”