Parallelized 0/1 Knapsack Algorithm Optimization in CPU-GPU-Based Heterogeneous System with Algorithm-based Fault Tolerance

A heterogeneous high-performance computing (HPC) system is an aggregation of CPUs and GPUs through high-speed interconnection. As graphics processing units (GPUs) can conduct rapid mathematical and vector operations due to their hardware properties, GPU-based heterogeneous systems can optimize algorithms that show data parallelism. Since the dynamic programming (DP) solution of the 0/1 knapsack algorithm is inherently data-parallel, this paper optimized the algorithm to maximize the advantages of a GPU-based parallel environment. Furthermore, this research identified that the algorithm was well suited to four distinct GPU-launch approaches. Among them, the “single-cell calculation by single thread with a single kernel launch” approach outperforms others by reducing the complexity from O(
$n$
*w) to O(
$n$
). Besides, these HPC systems are susceptible to the slightest of errors. Checkpoint recovery (CP/R) is the most commonly used fault tolerance and recovery system, which suffers from huge checkpoint and recovery overhead. To reduce this overhead in CPU-GPU-based heterogeneous systems, an algorithm-based fault tolerance (ABFT) technique has been proposed for a particular class of algorithm: 0/1 knapsack. The experimental results verify the feasibility of the proposed ABFT technique and demonstrate that the system can recover from failures with minimal overhead.